Mercurial Hosting > traffic-intelligence
view trafficintelligence/moving.py @ 1266:ebb18043616e
work in progress on categorization
author | Nicolas Saunier <nicolas.saunier@polymtl.ca> |
---|---|
date | Tue, 28 May 2024 17:16:41 -0400 |
parents | ef68d4ba7dae |
children | 20a5e1292321 |
line wrap: on
line source
#! /usr/bin/env python '''Libraries for moving objects, trajectories...''' import copy from math import sqrt, atan2, cos, sin, inf from numpy import median, mean, array, arange, zeros, ones, hypot, NaN, std, floor, ceil, float32, argwhere, minimum, issubdtype, integer as npinteger, percentile, full from matplotlib.pyplot import plot, text, arrow from scipy.spatial.distance import cdist from scipy.signal import savgol_filter try: from shapely.geometry import Polygon, Point as shapelyPoint from shapely.prepared import prep, PreparedGeometry shapelyAvailable = True except ImportError: print('Shapely library could not be loaded') shapelyAvailable = False from trafficintelligence import utils, cvutils from trafficintelligence.base import VideoFilenameAddable class Interval(object): '''Generic interval: a subset of real numbers (not iterable)''' def __init__(self, first=0, last=-1, revert = False): if revert and last<first: self.first=last self.last=first else: self.first=first self.last=last def __str__(self): return '{0}-{1}'.format(self.first, self.last) def __repr__(self): return self.__str__() def __eq__(self, other): return ((self.first == other.first) and (self.last == other.last)) or ((self.first == other.last) and (self.last == other.first)) def empty(self): return self.first > self.last def center(self): return (self.first+self.last)/2. def length(self): '''Returns the length of the interval''' return float(max(0,self.last-self.first)) def equal(self, i2): return self.first==i2.first and self.last == i2.last def getList(self): return [self.first, self.last] def contains(self, instant): return (self.first<=instant and self.last>=instant) def inside(self, interval2): '''Indicates if the temporal interval of self is comprised in interval2''' return (self.first >= interval2.first) and (self.last <= interval2.last) def shift(self, offset): self.first += offset self.last += offset @classmethod def parse(cls, s): if '-' in s: tmp = s.split('-') if len(tmp) == 2: return cls(int(tmp[0]), int(tmp[1])) # TODO with floats? print(s+' is not a valid representation of an interval') return None @classmethod def union(cls, interval1, interval2): '''Smallest interval comprising self and interval2''' return cls(min(interval1.first, interval2.first), max(interval1.last, interval2.last)) @classmethod def intersection(cls, interval1, interval2): '''Largest interval comprised in both self and interval2''' return cls(max(interval1.first, interval2.first), min(interval1.last, interval2.last)) def distance(self, interval2): if not Interval.intersection(self, interval2).empty(): return 0 elif self.first > interval2.last: return self.first - interval2.last elif self.last < interval2.first: return interval2.first - self.last else: return None @classmethod def unionIntervals(cls, intervals): 'returns the smallest interval containing all intervals' inter = cls(intervals[0].first, intervals[0].last) for i in intervals[1:]: inter = cls.union(inter, i) return inter class TimeInterval(Interval): '''Temporal interval: set of instants at fixed time step, between first and last, included For example: based on frame numbers (hence the modified length method) It may be modified directly by setting first and last It also (mostly) works with datetime.datetime''' def __init__(self, first=0, last=-1, revert = False): super(TimeInterval, self).__init__(first, last, revert) @staticmethod def fromInterval(inter): return TimeInterval(inter.first, inter.last) def __getitem__(self, i): if not self.empty(): if isinstance(i, int): return self.first+i else: raise TypeError("Invalid argument type.") #elif isinstance( key, slice ): def __iter__(self): self.iterInstantNum = -1 return self def __next__(self): if self.iterInstantNum >= self.length()-1: raise StopIteration else: self.iterInstantNum += 1 return self[self.iterInstantNum] def length(self): '''Returns the length of the interval''' return float(max(0,self.last-self.first+1)) def __len__(self): return self.length() # class BoundingPolygon: # '''Class for a polygon bounding a set of points # with methods to create intersection, unions... # ''' # We will use the polygon class of Shapely class STObject(object): '''Class for spatio-temporal object, i.e. with temporal and spatial existence (time interval and bounding polygon for positions (e.g. rectangle)). It may not mean that the object is defined for all time instants within the time interval''' def __init__(self, num = None, timeInterval = None, boundingPolygon = None): self.num = num self.timeInterval = timeInterval self.boundingPolygon = boundingPolygon def empty(self): return self.timeInterval.empty()# or not self.boudingPolygon def getNum(self): return self.num def __len__(self): return self.timeInterval.length() def length(self): return self.timeInterval.length() def getFirstInstant(self): return self.timeInterval.first def getLastInstant(self): return self.timeInterval.last def setFirstInstant(self, t): if t <= self.timeInterval.last: self.timeInterval.first = t else: print('new first instant is after last, not changing') def setLastInstant(self, t): if t >= self.timeInterval.first: self.timeInterval.last = t else: print('new last instant is before first, not changing') def getTimeInterval(self): return self.timeInterval def existsAtInstant(self, t): return self.timeInterval.contains(t) def commonTimeInterval(self, obj2): return TimeInterval.intersection(self.getTimeInterval(), obj2.getTimeInterval()) def shiftTimeInterval(self, offset): self.timeInterval.shift(offset) class Point(object): def __init__(self, x, y): self.x = x self.y = y def __str__(self): return '({:f},{:f})'.format(self.x,self.y) def __repr__(self): return self.__str__() def __eq__(self, other): return (self.x == other.x) and (self.y == other.y) def __add__(self, other): return Point(self.x+other.x, self.y+other.y) def __sub__(self, other): return Point(self.x-other.x, self.y-other.y) def __neg__(self): return Point(-self.x, -self.y) def __mul__(self, alpha): 'Warning, returns a new Point' return Point(self.x*alpha, self.y*alpha) def divide(self, alpha): 'Warning, returns a new Point' return Point(self.x/alpha, self.y/alpha) def __getitem__(self, i): if i == 0: return self.x elif i == 1: return self.y else: raise IndexError() def orthogonal(self, clockwise = True): 'Returns the orthogonal vector' if clockwise: return Point(self.y, -self.x) else: return Point(-self.y, self.x) def normalize(self): return self.divide(self.norm2()) def projectLocal(self, v, clockwise = True): 'Projects point projected on v, v.orthogonal()' e1 = v.normalize() e2 = e1.orthogonal(clockwise) return Point(Point.dot(self, e1), Point.dot(self, e2)) def rotate(self, theta): return Point(self.x*cos(theta)-self.y*sin(theta), self.x*sin(theta)+self.y*cos(theta)) def plot(self, options = 'o', **kwargs): plot([self.x], [self.y], options, **kwargs) @staticmethod def plotSegment(p1, p2, options = 'o', withOrigin = True, **kwargs): plot([p1.x, p2.x], [p1.y, p2.y], options, **kwargs) p1.plot('or', **kwargs) def angle(self): return atan2(self.y, self.x) def norm2Squared(self): '''2-norm distance (Euclidean distance)''' return self.x**2+self.y**2 def norm2(self): '''2-norm distance (Euclidean distance)''' return sqrt(self.norm2Squared()) def norm1(self): return abs(self.x)+abs(self.y) def normMax(self): return max(abs(self.x),abs(self.y)) def aslist(self): return [self.x, self.y] def astuple(self): return (self.x, self.y) def asint(self): return Point(int(round(self.x)), int(round(self.y))) if shapelyAvailable: def asShapely(self): return shapelyPoint(self.x, self.y) def homographyProject(self, homography): projected = cvutils.homographyProject(array([[self.x], [self.y]]), homography) return Point(projected[0], projected[1]) def inRectangle(self, xmin, xmax, ymin, ymax): return (xmin <= self.x <= xmax) and (ymin <= self.y <= ymax) def inPolygon(self, polygon): '''Indicates if the point x, y is inside the polygon (array of Nx2 coordinates of the polygon vertices) taken from http://www.ariel.com.au/a/python-point-int-poly.html Use Polygon.contains if Shapely is installed''' n = polygon.shape[0]; counter = 0; p1 = polygon[0,:]; for i in range(n+1): p2 = polygon[i % n,:]; if self.y > min(p1[1],p2[1]): if self.y <= max(p1[1],p2[1]): if self.x <= max(p1[0],p2[0]): if p1[1] != p2[1]: xinters = (self.y-p1[1])*(p2[0]-p1[0])/(p2[1]-p1[1])+p1[0]; if p1[0] == p2[0] or self.x <= xinters: counter+=1; p1=p2 return (counter%2 == 1); @staticmethod def fromList(p): return Point(p[0], p[1]) @staticmethod def dot(p1, p2): 'Scalar product' return p1.x*p2.x+p1.y*p2.y @staticmethod def cross(p1, p2): 'Cross product' return p1.x*p2.y-p1.y*p2.x @staticmethod def parallel(p1, p2): return Point.cross(p1, p2) == 0. @staticmethod def cosine(p1, p2): return Point.dot(p1,p2)/(p1.norm2()*p2.norm2()) @staticmethod def distanceNorm2(p1, p2): return (p1-p2).norm2() @staticmethod def plotAll(points, closePolygon = False, options = '', **kwargs): xCoords = [p.x for p in points] yCoords = [p.y for p in points] if closePolygon: xCoords.append(xCoords[0]) yCoords.append(yCoords[0]) plot(xCoords, yCoords, options, **kwargs) def similarOrientation(self, refDirection, cosineThreshold): 'Indicates whether the cosine of the vector and refDirection is smaller than cosineThreshold' return Point.cosine(self, refDirection) >= cosineThreshold @staticmethod def timeToCollision(p1, p2, v1, v2, collisionThreshold): '''Computes exact time to collision with a distance threshold The unknown of the equation is the time to reach the intersection between the relative trajectory of one road user and the circle of radius collisionThreshold around the other road user''' dv = v1-v2 dp = p1-p2 a = dv.norm2Squared()#(v1.x-v2.x)**2 + (v1.y-v2.y)**2 b = 2*Point.dot(dv, dp)#2 * ((p1.x-p2.x) * (v1.x-v2.x) + (p1.y-p2.y) * (v1.y-v2.y)) c = dp.norm2Squared() - collisionThreshold**2#(p1.x-p2.x)**2 + (p1.y-p2.y)**2 - collisionThreshold**2 delta = b**2 - 4*a*c if delta >= 0: deltaRoot = sqrt(delta) ttc1 = (-b + deltaRoot)/(2*a) ttc2 = (-b - deltaRoot)/(2*a) if ttc1 >= 0 and ttc2 >= 0: return min(ttc1,ttc2) elif ttc1 >= 0: return ttc1 elif ttc2 >= 0: return ttc2 else: # ttc1 < 0 and ttc2 < 0: return None else: return None @staticmethod def timeToCollisionPoly(corners1, v1, corners2, v2): """ :param bbox_1: list: [bbox_1 array 4*2 , v_1_x (float), v_1_y (float)] :param bbox_2: [bbox_2_x array 4*2, v_2_x (float), v_2_y (float)] :return: ttc """ from sympy import solve from sympy.abc import t def NewFourPoints(bbox, col_time, v_x, v_y): return [i + col_time * v_x for i in bbox[0]], [j + col_time * v_y for j in bbox[1]] v_1_x, v_1_y = v1.x, v1.y v_2_x, v_2_y = v2.x, v2.y x_bbox_1 = [p.x for p in corners1] y_bbox_1 = [p.y for p in corners1] bbox_1 = [x_bbox_1, y_bbox_1] x_bbox_2 = [p.x for p in corners2] y_bbox_2 = [p.y for p in corners2] bbox_2 = [x_bbox_2, y_bbox_2] t_total = [] line = [[0, 1], [1, 2], [2, 3], [3, 0]] for i in range(4): for j in range(4): p0_x, p0_y = x_bbox_2[i], y_bbox_2[i] p1_x, p1_y = x_bbox_1[line[j][0]], y_bbox_1[line[j][0]] p2_x, p2_y = x_bbox_1[line[j][1]], y_bbox_1[line[j][1]] p0_x_t = p0_x + v_2_x * t p0_y_t = p0_y + v_2_y * t p1_x_t = p1_x + v_1_x * t p1_y_t = p1_y + v_1_y * t p2_x_t = p2_x + v_1_x * t p2_y_t = p2_y + v_1_y * t f = (p2_x_t - p1_x_t) * (p0_y_t - p1_y_t) - (p0_x_t - p1_x_t) * (p2_y_t - p1_y_t) collision_t = solve(f, t, set=True, dict=True) if collision_t and collision_t[0][t] >= 0: collision_t = collision_t[0][t] #print(f'collision_t: {collision_t} >>>>>>>>>') x_bbox_1_new, y_bbox_1_new = NewFourPoints(bbox_1, collision_t, v_1_x, v_1_y) x_bbox_2_new, y_bbox_2_new = NewFourPoints(bbox_2, collision_t, v_2_x, v_2_y) p0_x, p0_y = x_bbox_2_new[i], y_bbox_2_new[i] p1_x, p1_y = x_bbox_1_new[line[j][0]], y_bbox_1_new[line[j][0]] p2_x, p2_y = x_bbox_1_new[line[j][1]], y_bbox_1_new[line[j][1]] min_x, max_x = min(p1_x, p2_x), max(p1_x, p2_x) min_y, max_y = min(p1_y, p2_y), max(p1_y, p2_y) if min_x == max_x: if min_y <= p0_y <= max_y: t_total.append(collision_t) elif min_y == max_y: if min_x <= p0_x <= max_x: t_total.append(collision_t) else: if min_x <= p0_x <= max_x or min_y <= p0_y <= max_y: t_total.append(collision_t) x_temps = x_bbox_2 x_bbox_2 = x_bbox_1 x_bbox_1 = x_temps # print(f'Change bounding box1 and box2 >>>>>>>>>>>>>>>>>>>\n\n\n') for i in range(4): for j in range(4): p0_x, p0_y = x_bbox_2[i], y_bbox_2[i] p1_x, p1_y = x_bbox_1[line[j][0]], y_bbox_1[line[j][0]] p2_x, p2_y = x_bbox_1[line[j][1]], y_bbox_1[line[j][1]] p0_x_t = p0_x + v_2_x * t p0_y_t = p0_y + v_2_y * t p1_x_t = p1_x + v_1_x * t p1_y_t = p1_y + v_1_y * t p2_x_t = p2_x + v_1_x * t p2_y_t = p2_y + v_1_y * t f = (p2_x_t - p1_x_t) * (p0_y_t - p1_y_t) - (p0_x_t - p1_x_t) * (p2_y_t - p1_y_t) collision_t = solve(f, t, set=True, dict=True) if collision_t and collision_t[0][t] >= 0: collision_t = collision_t[0][t] # print(f'collision_t: {collision_t} >>>>>>>>>') x_bbox_1_new, y_bbox_1_new = NewFourPoints(bbox_1, collision_t, v_1_x, v_1_y) x_bbox_2_new, y_bbox_2_new = NewFourPoints(bbox_2, collision_t, v_2_x, v_2_y) p0_x, p0_y = x_bbox_2_new[i], y_bbox_2_new[i] p1_x, p1_y = x_bbox_1_new[line[j][0]], y_bbox_1_new[line[j][0]] p2_x, p2_y = x_bbox_1_new[line[j][1]], y_bbox_1_new[line[j][1]] min_x, max_x = min(p1_x, p2_x), max(p1_x, p2_x) min_y, max_y = min(p1_y, p2_y), max(p1_y, p2_y) if min_x == max_x: if min_y <= p0_y <= max_y: t_total.append(collision_t) elif min_y == max_y: if min_x <= p0_x <= max_x: t_total.append(collision_t) else: if min_x <= p0_x <= max_x or min_y <= p0_y <= max_y: t_total.append(collision_t) if t_total: collision_t_min = min(t_total) # print(f'collision_time:') # print(collision_t_min) else: collision_t_min = None # print(f'No collision') return collision_t_min @staticmethod def midPoint(p1, p2): 'Returns the middle of the segment [p1, p2]' return Point(0.5*p1.x+0.5*p2.x, 0.5*p1.y+0.5*p2.y) @staticmethod def agg(points, aggFunc = mean): return Point(aggFunc([p.x for p in points]), aggFunc([p.y for p in points])) @staticmethod def boundingRectangle(points, v): '''Returns the bounding rectangle of the points, aligned on the vector v A list of points is returned: front left, front right, rear right, rear left''' e1 = v.normalize() e2 = e1.orthogonal(False) xCoords = [] yCoords = [] for p in points: xCoords.append(Point.dot(e1, p)) yCoords.append(Point.dot(e2, p)) xmin = min(xCoords) xmax = max(xCoords) ymin = min(yCoords) ymax = max(yCoords) frontLeft = Point(xmax, ymax) frontRight = Point(xmax, ymin) rearLeft = Point(xmin, ymax) rearRight = Point(xmin, ymin) originalE1 = Point(Point.dot(e1, Point(1,0)),Point.dot(e2, Point(1,0))) originalE2 = Point(Point.dot(e1, Point(0,1)),Point.dot(e2, Point(0,1))) return [Point(Point.dot(originalE1, p), Point.dot(originalE2, p)) for p in [frontLeft, frontRight, rearRight, rearLeft]] if shapelyAvailable: def pointsToShapely(points): 'Returns shapely polygon' return Polygon([p.astuple() for p in points]) def pointsInPolygon(points, polygon): '''Optimized tests of a series of points within (Shapely) polygon (not prepared)''' if type(polygon) == PreparedGeometry: prepared_polygon = polygon else: prepared_polygon = prep(polygon) return list(filter(prepared_polygon.contains, points)) # Functions for coordinate transformation # From Paul St-Aubin's PVA tools def prepareAlignments(alignments): '''Prepares alignments (list of splines, each typically represented as a Trajectory) - computes cumulative distances - approximates slope singularity by giving some slope roundoff (account for roundoff error)''' for alignment in alignments: alignment.computeCumulativeDistances() p1 = alignment[0] for i in range(len(alignment)-1): p2 = alignment[i+1] if(round(p1.x, 10) == round(p2.x, 10)): p2.x += 0.0000000001 if(round(p1.y, 10) == round(p2.y, 10)): p2.y += 0.0000000001 p1 = p2 def ppldb2p(qx,qy, p0x,p0y, p1x,p1y): ''' Point-projection (Q) on line defined by 2 points (P0,P1). http://cs.nyu.edu/~yap/classes/visual/03s/hw/h2/math.pdf ''' if(p0x == p1x and p0y == p1y): return None try: #Approximate slope singularity by giving some slope roundoff; account for roundoff error # if(round(p0x, 10) == round(p1x, 10)): # p1x += 0.0000000001 # if(round(p0y, 10) == round(p1y, 10)): # p1y += 0.0000000001 #make the calculation Y = (-(qx)*(p0y-p1y)-(qy*(p0y-p1y)**2)/(p0x-p1x)+p0x**2*(p0y-p1y)/(p0x-p1x)-p0x*p1x*(p0y-p1y)/(p0x-p1x)-p0y*(p0x-p1x))/(p1x-p0x-(p0y-p1y)**2/(p0x-p1x)) X = (-Y*(p1y-p0y)+qx*(p1x-p0x)+qy*(p1y-p0y))/(p1x-p0x) except ZeroDivisionError: print('Error: Division by zero in ppldb2p. Please report this error with the full traceback:') print('qx={0}, qy={1}, p0x={2}, p0y={3}, p1x={4}, p1y={5}...'.format(qx, qy, p0x, p0y, p1x, p1y)) import pdb; pdb.set_trace() return Point(X, Y) def getSYfromXY(p, alignments, goodEnoughAlignmentDistance = 0.5): ''' Snap a point p to its nearest subsegment of it's nearest alignment (from the list alignments). A alignment is a list of points (class Point), most likely a trajectory. Output: ======= [alignment index, subsegment leading point index, snapped point, subsegment distance, alignment distance, orthogonal point offset] or None ''' minOffsetY = float('inf') #For each alignment for alignmentIdx in range(len(alignments)): #For each alignment point index for alignment_p in range(len(alignments[alignmentIdx])-1): #Get closest point on alignment closestPoint = ppldb2p(p.x,p.y,alignments[alignmentIdx][alignment_p][0],alignments[alignmentIdx][alignment_p][1],alignments[alignmentIdx][alignment_p+1][0],alignments[alignmentIdx][alignment_p+1][1]) if closestPoint is None: print('Error: Alignment {0}, segment {1} has identical bounds and therefore is not a vector. Projection cannot continue.'.format(alignmentIdx, alignment_p)) return None # check if the projected point is in between the current segment of the alignment bounds if utils.inBetween(alignments[alignmentIdx][alignment_p][0], alignments[alignmentIdx][alignment_p+1][0], closestPoint.x) and utils.inBetween(alignments[alignmentIdx][alignment_p][1], alignments[alignmentIdx][alignment_p+1][1], closestPoint.y): offsetY = Point.distanceNorm2(closestPoint, p) if offsetY < minOffsetY: minOffsetY = offsetY snappedAlignmentIdx = alignmentIdx snappedAlignmentLeadingPoint = alignment_p snappedPoint = Point(closestPoint.x, closestPoint.y) #Jump loop if significantly close if offsetY < goodEnoughAlignmentDistance: break #Get sub-segment distance if minOffsetY != float('inf'): subsegmentDistance = Point.distanceNorm2(snappedPoint, alignments[snappedAlignmentIdx][snappedAlignmentLeadingPoint]) #Get cumulative alignment distance (total segment distance) alignmentDistanceS = alignments[snappedAlignmentIdx].getCumulativeDistance(snappedAlignmentLeadingPoint) + subsegmentDistance orthogonalAlignmentVector = (alignments[snappedAlignmentIdx][snappedAlignmentLeadingPoint+1]-alignments[snappedAlignmentIdx][snappedAlignmentLeadingPoint]).orthogonal() offsetVector = p-snappedPoint if Point.dot(orthogonalAlignmentVector, offsetVector) < 0: minOffsetY = -minOffsetY return [snappedAlignmentIdx, snappedAlignmentLeadingPoint, snappedPoint, subsegmentDistance, alignmentDistanceS, minOffsetY] else: print('Offset for point {} is infinite (check with prepareAlignments if some alignment segments are aligned with axes)'.format(p)) return None def getXYfromSY(s, y, alignmentNum, alignments): ''' Find X,Y coordinate from S,Y data. if mode = 0 : return Snapped X,Y if mode !=0 : return Real X,Y ''' alignment = alignments[alignmentNum] i = 1 while s > alignment.getCumulativeDistance(i) and i < len(alignment): i += 1 if i < len(alignment): d = s - alignment.getCumulativeDistance(i-1) # distance on subsegment #Get difference vector and then snap dv = alignment[i] - alignment[i-1] normalizedV = dv.normalize() #snapped = alignment[i-1] + normalizedV*d # snapped point coordinate along alignment # add offset finally orthoNormalizedV = normalizedV.orthogonal() return alignment[i-1] + normalizedV*d + orthoNormalizedV*y else: print('Curvilinear point {} is past the end of the alignement'.format((s, y, alignmentNum))) return None class NormAngle(object): '''Alternate encoding of a point, by its norm and orientation''' def __init__(self, norm, angle): self.norm = norm self.angle = angle @staticmethod def fromPoint(p): norm = p.norm2() if norm > 0: angle = p.angle() else: angle = 0. return NormAngle(norm, angle) def __add__(self, other): 'a norm cannot become negative' return NormAngle(max(self.norm+other.norm, 0), self.angle+other.angle) def getPoint(self): return Point(self.norm*cos(self.angle), self.norm*sin(self.angle)) def predictPositionNoLimit(nTimeSteps, initialPosition, initialVelocity, initialAcceleration = Point(0,0)): '''Predicts the position in nTimeSteps at constant speed/acceleration''' return initialVelocity + initialAcceleration.__mul__(nTimeSteps),initialPosition+initialVelocity.__mul__(nTimeSteps) + initialAcceleration.__mul__(nTimeSteps**2*0.5) def predictPosition(position, speedOrientation, control, maxSpeed = None): '''Predicts the position (moving.Point) at the next time step with given control input (deltaSpeed, deltaTheta) speedOrientation is the other encoding of velocity, (speed, orientation) speedOrientation and control are NormAngle''' predictedSpeedTheta = speedOrientation+control if maxSpeed is not None: predictedSpeedTheta.norm = min(predictedSpeedTheta.norm, maxSpeed) predictedPosition = position+predictedSpeedTheta.getPoint() return predictedPosition, predictedSpeedTheta class FlowVector(object): '''Class to represent 4-D flow vectors, ie a position and a velocity''' def __init__(self, position, velocity): 'position and velocity should be Point instances' self.position = position self.velocity = velocity def __add__(self, other): return FlowVector(self.position+other.position, self.velocity+other.velocity) def __mul__(self, alpha): return FlowVector(self.position.__mul__(alpha), self.velocity.__mul__(alpha)) def plot(self, options = '', **kwargs): plot([self.position.x, self.position.x+self.velocity.x], [self.position.y, self.position.y+self.velocity.y], options, **kwargs) self.position.plot(options+'x', **kwargs) @staticmethod def similar(f1, f2, maxDistance2, maxDeltavelocity2): return (f1.position-f2.position).norm2Squared()<maxDistance2 and (f1.velocity-f2.velocity).norm2Squared()<maxDeltavelocity2 def intersection(p1, p2, p3, p4): ''' Intersection point (x,y) of the segments [p1, p2] and [p3, p4] Returns the intersection point and the ratio of its position along [p1, p2] from p1 Based on http://paulbourke.net/geometry/pointlineplane/''' dp12 = p2-p1 dp34 = p4-p3 #det = (p4.y-p3.y)*(p2.x-p1.x)-(p4.x-p3.x)*(p2.y-p1.y) det = float(dp34.y*dp12.x-dp34.x*dp12.y) if det == 0.: return None, None else: ua = (dp34.x*(p1.y-p3.y)-dp34.y*(p1.x-p3.x))/det return p1+dp12.__mul__(ua), ua # def intersection(p1, p2, dp1, dp2): # '''Returns the intersection point between the two lines # defined by the respective vectors (dp) and origin points (p)''' # from numpy import matrix # from numpy.linalg import linalg # A = matrix([[dp1.y, -dp1.x], # [dp2.y, -dp2.x]]) # B = matrix([[dp1.y*p1.x-dp1.x*p1.y], # [dp2.y*p2.x-dp2.x*p2.y]]) # if linalg.det(A) == 0: # return None # else: # intersection = linalg.solve(A,B) # return Point(intersection[0,0], intersection[1,0]) def segmentIntersection(p1, p2, p3, p4): '''Returns the intersecting point (and ratio along [p1, p2]) of the segments [p1, p2] and [p3, p4], None otherwise''' if (Interval.intersection(Interval(p1.x,p2.x,True), Interval(p3.x,p4.x,True)).empty()) or (Interval.intersection(Interval(p1.y,p2.y,True), Interval(p3.y,p4.y,True)).empty()): return None, None else: inter, ratio = intersection(p1, p2, p3, p4) if (inter is not None and utils.inBetween(p1.x, p2.x, inter.x) and utils.inBetween(p3.x, p4.x, inter.x) and utils.inBetween(p1.y, p2.y, inter.y) and utils.inBetween(p3.y, p4.y, inter.y)): return inter, ratio else: return None, None def segmentLineIntersection(p1, p2, p3, p4): '''Indicates if the line going through p1 and p2 intersects inside p3, p4''' inter, ratio = intersection(p1, p2, p3, p4) if inter is not None and utils.inBetween(p3.x, p4.x, inter.x) and utils.inBetween(p3.y, p4.y, inter.y): return inter, ratio else: return None, None def segmentOrientationCrossing(p1, p2, p3, p4): '''Returns the direction of the crossing, assuming there is a crossing: True means right (p3) to left (p4) (along the orthogonal vector to [p1, p2] (positive trigonometric orientation), False the other way''' dp12 = p2-p1 ortho = dp12.orthogonal(False) return Point.dot(ortho, p4-p3) > 0 class Trajectory(object): '''Class for trajectories: temporal sequence of positions The class is iterable''' def __init__(self, positions=None): if positions is not None: self.positions = positions else: self.positions = [[],[]] @staticmethod def generate(p, v, nPoints): t = Trajectory() p0 = Point(p.x, p.y) t.addPosition(p0) for i in range(nPoints-1): p0 += v t.addPosition(p0) return t @staticmethod def load(line1, line2): return Trajectory([[float(n) for n in line1.split(' ')], [float(n) for n in line2.split(' ')]]) @staticmethod def fromPointList(points): t = Trajectory() if isinstance(points[0], list) or isinstance(points[0], tuple): for p in points: t.addPositionXY(p[0],p[1]) else: for p in points: t.addPosition(p) return t @staticmethod def fromPointDict(points): '''Points is a dict of points where keys are time instants and there are (probably) missing positions''' instants = sorted(list(points)) # find all gaps t1 = instants[0] gap = False gaps = [] for t in range(instants[0], instants[-1]+1): if t in instants: if gap: t2 = t # store gap gaps.append([t1, t2]) gap = False t1 = t else: t1 = t else: gap = True # interpolate for gaps for gap in gaps: v = (points[gap[1]]-points[gap[0]]).divide(gap[1]-gap[0]) for t in range(gap[0]+1, gap[1]): points[t]=points[t-1]+v return Trajectory.fromPointList([points[t] for t in range(instants[0], instants[-1]+1)]) def __len__(self): return len(self.positions[0]) def length(self): return self.__len__() def empty(self): return self.__len__() == 0 def __getitem__(self, i): if isinstance(i, int):# or issubdtype(i, npinteger): return Point(self.positions[0][i], self.positions[1][i]) elif isinstance(i, slice): return Trajectory([self.positions[0][i],self.positions[1][i]]) else: raise TypeError("Invalid argument type.") def __str__(self): return ' '.join([self.__getitem__(i).__str__() for i in range(self.length())]) def __repr__(self): return self.__str__() def __iter__(self): self.iterInstantNum = 0 return self def __next__(self): if self.iterInstantNum >= self.length(): raise StopIteration else: self.iterInstantNum += 1 return self[self.iterInstantNum-1] def __eq__(self, other): if self.length() == other.length(): result = True for p, po in zip(self, other): result = result and (p == po) return result else: return False def append(self,other): '''adds positions of other to the trajectory (in-place modification)''' for i in range(2): self.positions[i] += other.positions[i] def setPositionXY(self, i, x, y): if i < self.__len__(): self.positions[0][i] = x self.positions[1][i] = y def setPosition(self, i, p): self.setPositionXY(i, p.x, p.y) def reset(self, x, y): for i in range(self.__len__()): self.positions[0][i] = x self.positions[1][i] = y def addPositionXY(self, x, y): self.positions[0].append(x) self.positions[1].append(y) def addPosition(self, p): self.addPositionXY(p.x, p.y) def duplicateLastPosition(self): self.positions[0].append(self.positions[0][-1]) self.positions[1].append(self.positions[1][-1]) def agg(self, aggFunc = mean): return Point.agg(self, aggFunc) @staticmethod def _plot(positions, options = '', withOrigin = False, lastCoordinate = None, timeStep = 1, objNum = None, **kwargs): if lastCoordinate is None: plot(positions[0][::timeStep], positions[1][::timeStep], options, **kwargs) elif 0 <= lastCoordinate <= len(positions[0]): plot(positions[0][:lastCoordinate:timeStep], positions[1][:lastCoordinate:timeStep], options, **kwargs) if withOrigin: plot([positions[0][0]], [positions[1][0]], 'ro', **kwargs) if objNum is not None: text(positions[0][0], positions[1][0], '{}'.format(objNum)) def homographyProject(self, homography): return Trajectory(cvutils.homographyProject(array(self.positions), homography).tolist()) def newCameraProject(self, newCameraMatrix): return Trajectory(cvutils.newCameraProject(array(self.positions), newCameraMatrix).tolist()) def plot(self, options = '', withOrigin = False, timeStep = 1, objNum = None, **kwargs): Trajectory._plot(self.positions, options, withOrigin, None, timeStep, objNum, **kwargs) def plotAt(self, lastCoordinate, options = '', withOrigin = False, timeStep = 1, objNum = None, **kwargs): Trajectory._plot(self.positions, options, withOrigin, lastCoordinate, timeStep, objNum, **kwargs) def plotOnWorldImage(self, nPixelsPerUnitDistance, options = '', withOrigin = False, timeStep = 1, objNum = None, **kwargs): imgPositions = [[x*nPixelsPerUnitDistance for x in self.positions[0]], [x*nPixelsPerUnitDistance for x in self.positions[1]]] Trajectory._plot(imgPositions, options, withOrigin, None, timeStep, objNum, **kwargs) def getXCoordinates(self): return self.positions[0] def getYCoordinates(self): return self.positions[1] def asArray(self): return array(self.positions) def xBounds(self): # look for function that does min and max in one pass return Interval(min(self.getXCoordinates()), max(self.getXCoordinates())) def yBounds(self): # look for function that does min and max in one pass return Interval(min(self.getYCoordinates()), max(self.getYCoordinates())) def add(self, traj2): '''Returns a new trajectory of the same length''' if self.length() != traj2.length(): print('Trajectories of different lengths') return None else: return Trajectory([[a+b for a,b in zip(self.getXCoordinates(),traj2.getXCoordinates())], [a+b for a,b in zip(self.getYCoordinates(),traj2.getYCoordinates())]]) def subtract(self, traj2): '''Returns a new trajectory of the same length''' if self.length() != traj2.length(): print('Trajectories of different lengths') return None else: return Trajectory([[a-b for a,b in zip(self.getXCoordinates(),traj2.getXCoordinates())], [a-b for a,b in zip(self.getYCoordinates(),traj2.getYCoordinates())]]) def __mul__(self, alpha): '''Returns a new trajectory of the same length''' return Trajectory([[alpha*x for x in self.getXCoordinates()], [alpha*y for y in self.getYCoordinates()]]) def filterMovingWindow(self, halfWidth): '''Returns a new Trajectory obtained after the smoothing of the input by a moving average''' return Trajectory([utils.filterMovingWindow(self.positions[0], halfWidth), utils.filterMovingWindow(self.positions[1], halfWidth)]) def differentiate(self, doubleLastPosition = False): diff = Trajectory() for i in range(1, self.length()): diff.addPosition(self[i]-self[i-1]) if doubleLastPosition: diff.addPosition(diff[-1]) return diff def filterSG(self, window_length, polyorder, deriv=0, delta=1.0, axis=-1, mode='nearest', cval=0.0, nInstantsIgnoredAtEnds = 2): '''Filters the trajectory using the Savitsky Golay filter if deriv = 1, the method differentiates Warning: high order polynomials yield artefacts window_length : The length of the filter window (i.e. the number of coefficients). window_length must be a positive odd integer. polyorder : The order of the polynomial used to fit the samples. polyorder must be less than window_length. deriv : The order of the derivative to compute. This must be a nonnegative integer. The default is 0, which means to filter the data without differentiating. delta : The spacing of the samples to which the filter will be applied. This is only used if deriv > 0. Default is 1.0. axis : The axis of the array x along which the filter is to be applied. Default is -1. mode : Must be mirror, constant, nearest, wrap or interp. This determines the type of extension to use for the padded signal to which the filter is applied. When mode is constant, the padding value is given by cval. See the Notes for more details on mirror, constant, wrap, and nearest. When the interp mode is selected (the default), no extension is used. Instead, a degree polyorder polynomial is fit to the last window_length values of the edges, and this polynomial is used to evaluate the last window_length // 2 output values. cval : Value to fill past the edges of the input if mode is constant. Default is 0.0. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.savgol_filter.html#scipy.signal.savgol_filter''' filtered = savgol_filter(self.positions, window_length, polyorder, deriv, delta, axis, mode, cval) length = self.length() if nInstantsIgnoredAtEnds >=1: if nInstantsIgnoredAtEnds >= length/2: n = int(round(length/2))-1 else: n = nInstantsIgnoredAtEnds filtered = filtered[:,n:-n] return Trajectory(filtered.tolist()) def norm(self): '''Returns the list of the norms at each instant''' return hypot(self.positions[0], self.positions[1]) def computeCumulativeDistances(self): '''Computes the distance from each point to the next and the cumulative distance up to the point Can be accessed through getDistance(idx) and getCumulativeDistance(idx)''' self.distances = [] self.cumulativeDistances = [0.] p1 = self[0] cumulativeDistance = 0. for i in range(self.length()-1): p2 = self[i+1] self.distances.append(Point.distanceNorm2(p1,p2)) cumulativeDistance += self.distances[-1] self.cumulativeDistances.append(cumulativeDistance) p1 = p2 def getDistance(self,i): '''Return the distance between points i and i+1''' if i < self.length()-1: return self.distances[i] else: print('Index {} beyond trajectory length {}-1'.format(i, self.length())) def getCumulativeDistance(self, i): '''Returns the cumulative distance between the beginning and point i''' if i < self.length(): return self.cumulativeDistances[i] else: print('Index {} beyond trajectory length {}'.format(i, self.length())) def getTotalDistance(self): '''Returns the total distance (shorthand for cumulative distance [-1]''' return self.getCumulativeDistance(-1) def getMaxDistance(self, metric): 'Returns the maximum distance between points in the trajectory' positions = self.getPositions().asArray().T return cdist(positions, positions, metric = metric).max() def getClosestPoint(self, p1, maxDist2 = None): '''Returns the instant of the closest position in trajectory to p1 (and the point) if maxDist is not None, will check the distance is smaller TODO: could use cdist for different metrics''' distances2 = [] minDist2 = float('inf') i = -1 for p2 in self: distances2.append(Point.distanceNorm2(p1, p2)) if distances2[-1] < minDist2: minDist2 = distances2[-1] i = len(distances2)-1 if maxDist2 is not None and minDist2 < maxDist2: return None else: return i def similarOrientation(self, refDirection, cosineThreshold, minProportion = 0.5): '''Indicates whether the minProportion (<=1.) (eg half) of the trajectory elements (vectors for velocity) have a cosine with refDirection is smaller than cosineThreshold''' count = 0 lengthThreshold = float(self.length())*minProportion for p in self: if p.similarOrientation(refDirection, cosineThreshold): count += 1 return count >= lengthThreshold def wiggliness(self): straightDistance = Point.distanceNorm2(self.__getitem__(0),self.__getitem__(self.length()-1)) if straightDistance > 0: return self.getCumulativeDistance(self.length()-1)/float(straightDistance) else: return None def getIntersections(self, p1, p2, computeOrientations = False): '''Returns a list of the indices at which the trajectory intersects with the segment of extremities p1 and p2 Returns an empty list if there is no crossing''' indices = [] intersections = [] rightToLeftOrientations = [] for i in range(self.length()-1): q1=self.__getitem__(i) q2=self.__getitem__(i+1) p, ratio = segmentIntersection(p1, p2, q1, q2) if p is not None: # if q1.x != q2.x: # ratio = (p.x-q1.x)/(q2.x-q1.x) # elif q1.y != q2.y: # ratio = (p.y-q1.y)/(q2.y-q1.y) # else: # ratio = 0 indices.append(i+ratio) intersections.append(p) if computeOrientations: rightToLeftOrientations.append(segmentOrientationCrossing(p1, p2, q1, q2)) return indices, intersections, rightToLeftOrientations def getLineIntersections(self, p1, p2): '''Returns a list of the indices at which the trajectory intersects with the line going through p1 and p2 Returns an empty list if there is no crossing''' indices = [] intersections = [] for i in range(self.length()-1): q1=self.__getitem__(i) q2=self.__getitem__(i+1) p, ratio = segmentLineIntersection(q1, q2, p1, p2) if p is not None: # if q1.x != q2.x: # ratio = (p.x-q1.x)/(q2.x-q1.x) # elif q1.y != q2.y: # ratio = (p.y-q1.y)/(q2.y-q1.y) # else: # ratio = 0 indices.append(i+ratio) intersections.append(p) return indices, intersections def subTrajectoryInInterval(self, inter): 'Returns all position between index inter.first and index.last (included)' if inter.first >=0 and inter.last<= self.length(): return Trajectory([self.positions[0][inter.first:inter.last+1], self.positions[1][inter.first:inter.last+1]]) else: return None def subSample(self, step): 'Returns the positions very step' return Trajectory([self.positions[0][::step], self.positions[1][::step]]) if shapelyAvailable: def getInstantsInPolygon(self, polygon): '''Returns the list of instants at which the trajectory is in the polygon''' instants = [] n = self.length() for t, x, y in zip(range(n), self.positions[0], self.positions[1]): if polygon.contains(shapelyPoint(x, y)): instants.append(t) return instants def getTrajectoryInPolygon(self, polygon, t2 = None): '''Returns the trajectory built with the set of points inside the (shapely) polygon The polygon could be a prepared polygon (faster) from prepared.prep t2 is another trajectory (could be velocities) which is filtered based on the first (self) trajectory''' traj = Trajectory() inPolygon = [] for x, y in zip(self.positions[0], self.positions[1]): inPolygon.append(polygon.contains(shapelyPoint(x, y))) if inPolygon[-1]: traj.addPositionXY(x, y) traj2 = Trajectory() if t2 is not None: for inp, x, y in zip(inPolygon, t2.positions[0], t2.positions[1]): if inp: traj2.addPositionXY(x, y) return traj, traj2 def proportionInPolygon(self, polygon, minProportion = 0.5): instants = self.getInstantsInPolygon(polygon) lengthThreshold = float(self.length())*minProportion return len(instants) >= lengthThreshold else: def getTrajectoryInPolygon(self, polygon, t2 = None): '''Returns the trajectory built with the set of points inside the polygon (array of Nx2 coordinates of the polygon vertices)''' traj = Trajectory() inPolygon = [] for p in self: inPolygon.append(p.inPolygon(polygon)) if inPolygon[-1]: traj.addPosition(p) traj2 = Trajectory() if t2 is not None: for inp, x, y in zip(inPolygon, t2.positions[0], t2.positions[1]): if inp: traj2.addPositionXY(p.x, p.y) return traj, traj2 def proportionInPolygon(self, polygon, minProportion = 0.5): inPolygon = [p.inPolygon(polygon) for p in self] lengthThreshold = float(self.length())*minProportion return sum(inPolygon) >= lengthThreshold @staticmethod def lcss(t1, t2, lcss): return lcss.compute(t1, t2) class CurvilinearTrajectory(Trajectory): '''Sub class of trajectory for trajectories with curvilinear coordinates and lane assignements longitudinal coordinate is stored as first coordinate (exterior name S) lateral coordinate is stored as second coordinate the third "lane" coordinate is for an alignment id, whether explicit for a list/dict of alignments, or implicit for a road with lane numbers''' def __init__(self, S = None, Y = None, lanes = None): if S is None or Y is None or len(S) != len(Y): self.positions = [[],[]] if S is not None and Y is not None and len(S) != len(Y): print("S and Y coordinates of different lengths\nInitializing to empty lists") else: self.positions = [S,Y] if lanes is None or len(lanes) != self.length(): self.lanes = [None]*int(self.length()) else: self.lanes = lanes @staticmethod def generate(s, v, nPoints, lane, y = 0): '''s is initial position, v is velocity 0 in lateral coordinate by default TODO 2D velocity for lane change?''' S = [s] for i in range(nPoints-1): S.append(S[-1]+v) Y = [y]*nPoints lanes = [lane]*nPoints return CurvilinearTrajectory(S, Y, lanes) @staticmethod def fromTrajectoryProjection(t, alignments, halfWidth = 3): ''' Add, for every object position, the class 'moving.CurvilinearTrajectory()' (curvilinearPositions instance) which holds information about the curvilinear coordinates using alignment metadata. From Paul St-Aubin's PVA tools ====== Input: ====== alignments = a list of alignments, where each alignment is a list of points (class Point). halfWidth = moving average window (in points) in which to smooth lane changes. As per tools_math.cat_mvgavg(), this term is a search *radius* around the center of the window. ''' curvilinearPositions = CurvilinearTrajectory() #For each point for i in range(int(t.length())): result = getSYfromXY(t[i], alignments) # Error handling if(result is None): print('Warning: trajectory at point {} {} has alignment errors (alignment snapping)\nCurvilinear trajectory could not be computed'.format(i, t[i])) else: [align, alignPoint, snappedPoint, subsegmentDistance, S, Y] = result curvilinearPositions.addPositionSYL(S, Y, align) ## Go back through points and correct lane #Run through objects looking for outlier point smoothed_lanes = utils.filterCategoricalMovingWindow(curvilinearPositions.getLanes(), halfWidth) ## Recalculate projected point to new lane lanes = curvilinearPositions.getLanes() if(lanes != smoothed_lanes): for i in range(len(lanes)): if(lanes[i] != smoothed_lanes[i]): result = getSYfromXY(t[i],[alignments[smoothed_lanes[i]]]) # Error handling if(result is None): ## This can be triggered by tracking errors when the trajectory jumps around passed another alignment. print(' Warning: trajectory at point {} {} has alignment errors during trajectory smoothing and will not be corrected.'.format(i, t[i])) else: [align, alignPoint, snappedPoint, subsegmentDistance, S, Y] = result curvilinearPositions.setPosition(i, S, Y, align) return curvilinearPositions def __getitem__(self,i): if isinstance(i, int): return [self.positions[0][i], self.positions[1][i], self.lanes[i]] else: raise TypeError("Invalid argument type.") #elif isinstance( key, slice ): def getSCoordinates(self): return self.getXCoordinates() def getLanes(self): return self.lanes def getSCoordAt(self, i): return self.positions[0][i] def getYCoordAt(self, i): return self.positions[1][i] def getLaneAt(self, i): return self.lanes[i] def subTrajectoryInInterval(self, inter): 'Returns all curvilinear positions between index inter.first and index.last (included)' if inter.first >=0 and inter.last<= self.length(): return CurvilinearTrajectory(self.positions[0][inter.first:inter.last+1], self.positions[1][inter.first:inter.last+1], self.lanes[inter.first:inter.last+1]) else: return None def append(self, other): Trajectory.append(self, other) self.lanes.append(other.getLanes()) def addPositionSYL(self, s, y, lane = None): self.addPositionXY(s,y) self.lanes.append(lane) def addPosition(self, p): 'Adds position in the point format for curvilinear of list with 3 values' self.addPositionSYL(p[0], p[1], p[2]) def duplicateLastPosition(self): super(CurvilinearTrajectory, self).duplicateLastPosition() self.lanes.append(self.lanes[-1]) def setPosition(self, i, s, y, lane): self.setPositionXY(i, s, y) if i < self.__len__(): self.lanes[i] = lane def differentiate(self, doubleLastPosition = False): diff = CurvilinearTrajectory() p1 = self[0] for i in range(1, self.length()): p2 = self[i] if p2[2] == p1[2]: laneChange = None else: laneChange = (p1[2], p2[2]) diff.addPositionSYL(p2[0]-p1[0], p2[1]-p1[1], laneChange) p1=p2 if doubleLastPosition and self.length() > 1: diff.addPosition(diff[-1]) return diff def getIntersections(self, S1, lane = None): '''Returns a list of the indices at which the trajectory goes past the curvilinear coordinate S1 (in provided lane if lane is not None) Returns an empty list if there is no crossing''' indices = [] for i in range(self.length()-1): q1=self.__getitem__(i) q2=self.__getitem__(i+1) if q1[0] <= S1 < q2[0] and (lane is None or (self.lanes[i] == lane and self.lanes[i+1] == lane)): indices.append(i+(S1-q1[0])/(q2[0]-q1[0])) return indices ################## # Geometry ################## class Geometry(object): '''Generic class for outline and size of object ''' class CarGeometry(Geometry): '''Default car geometry as rectangle ''' ################## # Moving Objects ################## userTypeNames = ['unknown', 'car', 'pedestrian', 'motorcyclist', 'cyclist', 'bus', 'truck', 'automated'] coco2Types = {0: 2, 1: 4, 2: 1, 3: 3, 5: 5, 7: 6} cocoTypeNames = {0: 'person', 1: 'bicycle', 2: 'car', 3: 'motorcycle', 5: 'bus', #6: 'train', 7: 'truck'} kitti2Types = {'Car': 1, 'Cyclist': 4, 'Pedestrian': 2, 'Van': 1} userType2Num = utils.inverseEnumeration(userTypeNames) class CarClassifier: def predict(self, hog): return userType2Num['car'] carClassifier = CarClassifier() class MovingObject(STObject, VideoFilenameAddable): '''Class for moving objects: a spatio-temporal object with a trajectory and a geometry (constant volume over time) and a usertype (e.g. road user) coded as a number (see userTypeNames) ''' def __init__(self, num = None, timeInterval = None, positions = None, velocities = None, geometry = None, userType = userType2Num['unknown'], nObjects = None, features = None, initCurvilinear = False): super(MovingObject, self).__init__(num, timeInterval) if initCurvilinear: self.curvilinearPositions = positions self.curvilinearVelocities = velocities # third component is (previousAlignmentIdx, newAlignmentIdx) or None if no change else: self.positions = positions self.velocities = velocities self.geometry = geometry self.userType = userType self.setNObjects(nObjects) # a feature has None for nObjects self.features = features if features is not None and len(features) > 1: self.featureNumbers = [f.getNum() for f in features] # compute bounding polygon from trajectory @staticmethod def croppedTimeInterval(obj, value, after = True): newTimeInterval = TimeInterval(obj.getFirstInstant(), min(value, obj.getLastInstant())) if after else TimeInterval(max(obj.getFirstInstant(), value), obj.getLastInstant()) if obj.positions is not None : newPositions = obj.positions[slice(newTimeInterval.first - obj.getLastInstant(), newTimeInterval.last + 1 - obj.getLastInstant())] else: newPositions = None if obj.velocities is not None : newVelocities = obj.velocities[slice(newTimeInterval.first - obj.getLastInstant(), newTimeInterval.last + 1 - obj.getLastInstant())] else: newVelocities = None if obj.hasFeatures(): newFeatures = [f.croppedTimeInterval(value, after) for f in obj.features] else: newFeatures = None res = MovingObject(obj.getNum(), newTimeInterval, newPositions, newVelocities, obj.geometry, obj.userType, obj.nObjects) res.features = newFeatures res.featureNumbers = obj.featureNumbers #if hasattr(obj, 'projectedPositions'): # res.projectedPositions = obj.projectedPositions[slice(newTimeInterval.first, newTimeInterval.last+1)] return res @staticmethod def aggregateTrajectories(features, interval = None, aggFunc = mean): 'Computes the aggregate trajectory from list of MovingObject features' positions = Trajectory() velocities = Trajectory() if interval is None: inter = TimeInterval.unionIntervals([f.getTimeInterval() for f in features]) else: inter = interval for t in inter: points = [] vels = [] for f in features: if f.existsAtInstant(t): points.append(f.getPositionAtInstant(t)) vels.append(f.getVelocityAtInstant(t)) positions.addPosition(Point.agg(points, aggFunc)) velocities.addPosition(Point.agg(vels, aggFunc)) return inter, positions, velocities @staticmethod def generate(num, p, v, timeInterval): nPoints = int(timeInterval.length()) positions = Trajectory.generate(p, v, nPoints) return MovingObject(num = num, timeInterval = timeInterval, positions = positions, velocities = Trajectory([[v.x]*nPoints, [v.y]*nPoints])) def updatePositions(self): inter, self.positions, self.velocities = MovingObject.aggregateTrajectories(self.features, self.getTimeInterval()) @staticmethod def concatenate(obj1, obj2, num = None, newFeatureNum = None): '''Concatenates two objects, whether overlapping temporally or not Positions will be recomputed if features are merged Otherwise, only featureNumbers and/or features will be merged''' if num is None: newNum = obj1.getNum() else: newNum = num commonTimeInterval = obj1.commonTimeInterval(obj2) if commonTimeInterval.empty(): # and emptyInterval.length() >= 2: not needed 2 to include situations successive positions, but issues with missing last velocity and reloading emptyInterval = TimeInterval(min(obj1.getLastInstant(),obj2.getLastInstant()), max(obj1.getFirstInstant(),obj2.getFirstInstant())) if newFeatureNum is None: print('Not merging objects {} and {}, missing new feature number'.format(obj1.getNum(),obj2.getNum())) return None, None else: if obj1.existsAtInstant(emptyInterval.last): firstObject = obj2 secondObject = obj1 else: firstObject = obj1 secondObject = obj2 v = (secondObject.getPositionAtInstant(emptyInterval.last)-firstObject.getPositionAtInstant(emptyInterval.first)).divide(emptyInterval.length()-1) positions = copy.deepcopy(firstObject.getPositions()) velocities = copy.deepcopy(firstObject.getPositions()) featurePositions = Trajectory() featureVelocities = Trajectory() p = firstObject.getPositionAtInstant(emptyInterval.first) # init new feature with position at last instant of 1st obj featurePositions.addPosition(p) featureVelocities.addPosition(v) for t in range(emptyInterval.first+1, emptyInterval.last): p=p+v positions.addPosition(p) velocities.addPosition(v) featurePositions.addPosition(p) featureVelocities.addPosition(v) # last position to feature p = secondObject.getPositionAtInstant(emptyInterval.last) v = secondObject.getVelocityAtInstant(emptyInterval.last) featurePositions.addPosition(p) featureVelocities.addPosition(v) # copy second trajectory positions.append(secondObject.getPositions()) velocities.append(secondObject.getVelocities()) newObject = MovingObject(newNum, TimeInterval(firstObject.getFirstInstant(), secondObject.getLastInstant()), positions, velocities, nObjects = 1) newFeature = MovingObject(newFeatureNum, emptyInterval, featurePositions, featureVelocities) if hasattr(obj1, 'featureNumbers') and hasattr(obj2, 'featureNumbers'): newObject.featureNumbers = obj1.featureNumbers+obj2.featureNumbers+[newFeatureNum] if obj1.hasFeatures() and obj2.hasFeatures(): newObject.features = obj1.getFeatures()+obj2.getFeatures()+[newFeature] else: # time intervals overlap newTimeInterval = TimeInterval.union(obj1.getTimeInterval(), obj2.getTimeInterval()) newObject = MovingObject(newNum, newTimeInterval, nObjects = 1) # hypothesis is that it's the same object being reunited newFeature = None if hasattr(obj1, 'featureNumbers') and hasattr(obj2, 'featureNumbers'): newObject.featureNumbers = obj1.featureNumbers+obj2.featureNumbers if obj1.hasFeatures() and obj2.hasFeatures(): newObject.features = obj1.getFeatures()+obj2.getFeatures() newObject.updatePositions() else: print('Cannot update object positions without features') # user type if obj1.getUserType() != obj2.getUserType(): print('The two moving objects have different user types: obj1 {} obj2 {}'.format(userTypeNames[obj1.getUserType()], userTypeNames[obj2.getUserType()])) newObject.setUserType(obj1.getUserType()) return newObject, newFeature def getObjectInTimeInterval(self, inter): '''Returns a new object extracted from self, restricted to time interval inter''' intersection = TimeInterval.intersection(inter, self.getTimeInterval()) if not intersection.empty(): trajectoryInterval = TimeInterval(intersection.first-self.getFirstInstant(), intersection.last-self.getFirstInstant()) obj = MovingObject(self.num, intersection, self.positions.subTrajectoryInInterval(trajectoryInterval), self.geometry, self.userType, self.nObjects) if self.velocities is not None: obj.velocities = self.velocities.subTrajectoryInInterval(trajectoryInterval) return obj else: print('The object does not exist at {}'.format(inter)) return None def getObjectsInMask(self, mask, homography = None, minLength = 1): '''Returns new objects made of the positions in the mask mask is in the destination of the homography space''' if homography is not None: self.projectedPositions = self.positions.homographyProject(homography) else: self.projectedPositions = self.positions def inMask(positions, i, mask): p = positions[i] return mask[int(p.y), int(p.x)] != 0. #subTimeIntervals self.getFirstInstant()+i filteredIndices = [inMask(self.projectedPositions, i, mask) for i in range(int(self.length()))] # 'connected components' in subTimeIntervals l = 0 intervalLabels = [] prev = True for i in filteredIndices: if i: if not prev: # new interval l += 1 intervalLabels.append(l) else: intervalLabels.append(-1) prev = i intervalLabels = array(intervalLabels) subObjects = [] for l in set(intervalLabels): if l >= 0: if sum(intervalLabels == l) >= minLength: times = [self.getFirstInstant()+i for i in range(len(intervalLabels)) if intervalLabels[i] == l] subTimeInterval = TimeInterval(min(times), max(times)) subObjects.append(self.getObjectInTimeInterval(subTimeInterval)) return subObjects def getPositions(self): return self.positions def getVelocities(self): return self.velocities def getUserType(self): return self.userType def computeCumulativeDistances(self): self.positions.computeCumulativeDistances() def getCurvilinearPositions(self): if hasattr(self, 'curvilinearPositions'): return self.curvilinearPositions else: return None def getCurvilinearVelocities(self): if hasattr(self, 'curvilinearVelocities'): return self.curvilinearVelocities else: return None def plotCurvilinearPositions(self, lane = None, options = '', withOrigin = False, **kwargs): if hasattr(self, 'curvilinearPositions'): if lane is None: plot(list(self.getTimeInterval()), self.curvilinearPositions.positions[0], options, **kwargs) if withOrigin: plot([self.getFirstInstant()], [self.curvilinearPositions.positions[0][0]], 'ro', **kwargs) else: instants = [] coords = [] for t, p in zip(self.getTimeInterval(), self.curvilinearPositions): if p[2] == lane: instants.append(t) coords.append(p[0]) else: instants.append(NaN) coords.append(NaN) plot(instants, coords, options, **kwargs) if withOrigin and len(instants)>0: plot([instants[0]], [coords[0]], 'ro', **kwargs) else: print('Object {} has no curvilinear positions'.format(self.getNum())) def interpolateCurvilinearPositions(self, t, alignments = None): '''Linear interpolation of curvilinear positions, t being a float''' if hasattr(self, 'curvilinearPositions'): if self.existsAtInstant(t): i = int(floor(t)) p1 = self.getCurvilinearPositionAtInstant(i) p2 = self.getCurvilinearPositionAtInstant(i+1) if p1[2] == p2[2]: alpha = t-float(i) return [(1-alpha)*p1[0]+alpha*p2[0], (1-alpha)*p1[1]+alpha*p2[1], p1[2]] elif alignments is not None: # can be done only if assuming there is no missing alignmentn where the object has no coordinate pass # TODO else: print('Object {} changes lane at {} and alignments are not provided'.format(self.getNum(), t)) else: print('Object {} does not exist at {}'.format(self.getNum(), t)) else: print('Object {} has no curvilinear positions'.format(self.getNum())) def setUserType(self, userType): self.userType = userType def getNObjects(self): return self.nObjects def setNObjects(self, nObjects): if nObjects is None or nObjects >= 1: self.nObjects = nObjects else: print('Number of objects represented by object {} must be greater or equal to 1 ({})'.format(self.getNum(), nObjects)) self.nObjects = None def setFeatures(self, features, featuresOrdered = False): '''Sets the features in the features field based on featureNumbers if not all features are loaded from 0, one needs to renumber in a dict''' if featuresOrdered: tmp = features else: tmp = {f.getNum():f for f in features} self.features = [tmp[i] for i in self.featureNumbers] def getFeatures(self): return self.features def hasFeatures(self): return (self.features is not None) def getFeature(self, i): if self.hasFeatures() and i<len(self.features): return self.features[i] else: return None def getNLongestFeatures(self, nFeatures = 1): if self.features is None: return [] else: tmp = utils.sortByLength(self.getFeatures(), reverse = True) return tmp[:min(len(tmp), nFeatures)] def getFeatureNumbersOverTime(self): '''Returns the number of features at each instant dict instant -> number of features''' if self.hasFeatures(): featureNumbers = {} for t in self.getTimeInterval(): n = 0 for f in self.getFeatures(): if f.existsAtInstant(t): n += 1 featureNumbers[t]=n return featureNumbers else: print('Object {} has no features loaded.'.format(self.getNum())) return None def getSpeeds(self, nInstantsIgnoredAtEnds = 0): '''Returns the scalar speed''' speeds = self.getVelocities().norm() if nInstantsIgnoredAtEnds > 0: n = min(nInstantsIgnoredAtEnds, int(floor(self.length()/2.))) return speeds[n:-n] else: return speeds def getAccelerations(self, fromSpeeds = True, nInstantsIgnoredAtEnds = 0): '''Returns the scalar acceleration by differentiating the speeds (fromSpeeds = True) or by differentiating the velocity (vector) and taking the norm (frompSpeeds = False. In the second case, speed is always positive''' if fromSpeeds: speeds = array(self.getSpeeds()) accelerations = speeds[1:]-speeds[:-1] else: accelerations = self.getVelocities().differentiate().norm() if nInstantsIgnoredAtEnds > 0: n = min(nInstantsIgnoredAtEnds, int(floor(len(accelerations)/2.))) return accelerations[n:-n] else: return accelerations def getAccelerationsSG(self, window_length, polyorder, delta=1.0, axis=-1, mode='nearest', cval=0.0, nInstantsIgnoredAtEnds = 0): '''Returns the 1-D acceleration from the 1-D speeds Caution about previously filtered data''' speeds = self.getSpeeds() wlength = min(window_length, len(speeds)) if wlength % 2 == 0: wlength -=1 filtered = savgol_filter(speeds, wlength, min(wlength-1, polyorder), 1, delta, axis, mode, cval) if nInstantsIgnoredAtEnds >= 1: if nInstantsIgnoredAtEnds >= len(speeds)/2: n = int(round(len(speeds)/2))-1 else: n = nInstantsIgnoredAtEnds return filtered[n:-n] else: return filtered def getSpeedIndicator(self): from indicators import SeverityIndicator return SeverityIndicator('Speed', {t:self.getVelocityAtInstant(t).norm2() for t in self.getTimeInterval()}) def getPositionAt(self, i): return self.positions[i] def getVelocityAt(self, i): return self.velocities[i] def getPositionAtInstant(self, i): return self.positions[i-self.getFirstInstant()] def getVelocityAtInstant(self, i): return self.velocities[i-self.getFirstInstant()] def getCurvilinearPositionAt(self, i): return self.curvilinearPositions[i] def getCurvilinearVelocityAt(self, i): return self.curvilinearVelocities[i] def getCurvilinearPositionAtInstant(self, i): return self.curvilinearPositions[i-self.getFirstInstant()] def getCurvilinearVelocityAtInstant(self, i): return self.curvilinearVelocities[i-self.getFirstInstant()] def getXCoordinates(self): return self.positions.getXCoordinates() def getYCoordinates(self): return self.positions.getYCoordinates() def setStationary(self): '''Resets the position to the mean of existing positions and sets speeds to 0 And does the same to the features TODO: other options (provide x, y, what to do with features?)''' meanPosition = self.positions.agg(mean) self.positions.reset(meanPosition.x, meanPosition.y) self.velocities.reset(0,0) if self.hasFeatures(): for f in self.getFeatures(): f.setStationary() def plot(self, options = '', withOrigin = False, timeStep = 1, withFeatures = False, withOutline = False, withIds = False, **kwargs): if withIds: objNum = self.getNum() else: objNum = None if (withFeatures or withOutline) and self.hasFeatures(): if withOutline: for t in self.getTimeInterval(): if t%timeStep == 0: self.plotOutlineAtInstant(t, cvutils.colors[self.getNum()%len(cvutils.colors)]+'-') self.positions.plot(cvutils.colors[self.getNum()%len(cvutils.colors)]+'-', withOrigin, timeStep, objNum, **kwargs) else: for f in self.getFeatures(): f.positions.plot('r', True, timeStep, **kwargs) self.positions.plot('bx-', True, timeStep, objNum, **kwargs) else: self.positions.plot(options, withOrigin, timeStep, objNum, **kwargs) def plotOnWorldImage(self, nPixelsPerUnitDistance, options = '', withOrigin = False, timeStep = 1, withIds = False, **kwargs): if withIds: self.positions.plotOnWorldImage(nPixelsPerUnitDistance, options, withOrigin, timeStep, self.getNum(), **kwargs) else: self.positions.plotOnWorldImage(nPixelsPerUnitDistance, options, withOrigin, timeStep, None, **kwargs) def plotOutlineAtInstant(self, t, options = '', withVelocity = False, velocityMultiply = 5, arrowWidth=0.1, **kwargs): if self.hasFeatures(): points = [f.getPositionAtInstant(t) for f in self.getFeatures()] Point.plotAll(points, True, options, **kwargs) if withVelocity: p = self.getPositionAtInstant(t) v = self.getVelocityAtInstant(t)*velocityMultiply arrow(p.x, p.y, v.x, v.y, width=arrowWidth) def play(self, videoFilename, homography = None, undistort = False, intrinsicCameraMatrix = None, distortionCoefficients = None, undistortedImageMultiplication = 1.): cvutils.displayTrajectories(videoFilename, [self], homography = homography, firstFrameNum = self.getFirstInstant(), lastFrameNum = self.getLastInstant(), undistort = undistort, intrinsicCameraMatrix = intrinsicCameraMatrix, distortionCoefficients = distortionCoefficients, undistortedImageMultiplication = undistortedImageMultiplication) def speedDiagnostics(self, framerate = 1., display = False, nInstantsIgnoredAtEnds=0): speeds = framerate*self.getSpeeds(nInstantsIgnoredAtEnds) coef = utils.linearRegression(list(range(len(speeds))), speeds) print('min/5th perc speed: {} / {}\nspeed diff: {}\nspeed stdev: {}\nregression: {}'.format(min(speeds), percentile(speeds, 5), speeds[-2]-speeds[1], std(speeds), coef[0])) if display: from matplotlib.pyplot import figure, axis figure(1) self.plot() axis('equal') figure(2) plot(list(self.getTimeInterval()), speeds) figure(3) plot(list(self.getTimeInterval()), self.getAccelerations(9, 3, speeds = speeds)) # arbitrary parameter @staticmethod def minMaxDistance(obj1, obj2): '''Computes the min max distance used for feature grouping''' commonTimeInterval = obj1.commonTimeInterval(obj2) if not commonTimeInterval.empty(): minDistance = (obj1.getPositionAtInstant(commonTimeInterval.first)-obj2.getPositionAtInstant(commonTimeInterval.first)).norm2() maxDistance = minDistance for t in list(commonTimeInterval)[1:]: d = (obj1.getPositionAtInstant(t)-obj2.getPositionAtInstant(t)).norm2() if d<minDistance: minDistance = d elif d>maxDistance: maxDistance = d return int(commonTimeInterval.length()), minDistance, maxDistance else: return int(commonTimeInterval.length()), None, None @staticmethod def distances(obj1, obj2, instant1, _instant2 = None): '''Returns the distances between all features of the 2 objects at the same instant instant1 or at instant1 and instant2''' if _instant2 is None: instant2 = instant1 else: instant2 = _instant2 positions1 = [f.getPositionAtInstant(instant1).astuple() for f in obj1.features if f.existsAtInstant(instant1)] positions2 = [f.getPositionAtInstant(instant2).astuple() for f in obj2.features if f.existsAtInstant(instant2)] return cdist(positions1, positions2, metric = 'euclidean') @staticmethod def minDistance(obj1, obj2, instant1, instant2 = None): return MovingObject.distances(obj1, obj2, instant1, instant2).min() @staticmethod def maxDistance(obj1, obj2, instant, instant2 = None): return MovingObject.distances(obj1, obj2, instant1, instant2).max() def maxSize(self): '''Returns the max distance between features at instant there are the most features''' if hasattr(self, 'features'): nFeatures = -1 tMaxFeatures = 0 for t in self.getTimeInterval(): n = len([f for f in self.features if f.existsAtInstant(t)]) if n > nFeatures: nFeatures = n tMaxFeatures = t return MovingObject.maxDistance(self, self, tMaxFeatures) else: print('Load features to compute a maximum size') return None def setRoutes(self, startRouteID, endRouteID): self.startRouteID = startRouteID self.endRouteID = endRouteID def getInstantsCrossingLine(self, p1, p2, computeOrientations = False): '''Returns the instant(s) at which the object passes from one side of the segment to the other empty list if there is no crossing''' indices, intersections, rightToLeftOrientations = self.positions.getIntersections(p1, p2, computeOrientations) return [t+self.getFirstInstant() for t in indices], intersections, rightToLeftOrientations def computeTrajectorySimilarities(self, prototypes, lcss): 'Computes the similarities to the prototypes using the LCSS' if not hasattr(self, 'prototypeSimilarities'): self.prototypeSimilarities = [] for proto in prototypes: lcss.similarities(proto.getMovingObject().getPositions().asArray().T, self.getPositions().asArray().T) similarities = lcss.similarityTable[-1, :-1].astype(float) self.prototypeSimilarities.append(similarities/minimum(arange(1., len(similarities)+1), proto.getMovingObject().length()*ones(len(similarities)))) @staticmethod def computePET(obj1, obj2, collisionDistanceThreshold = None, useBoundingPoly = False): '''Post-encroachment time based on distance threshold Returns the smallest time difference when the object positions are within collisionDistanceThreshold and the instants at which each object is passing through its corresponding position''' n1 = int(obj1.length()) n2 = int(obj2.length()) pets = zeros((n1, n2)) for i,t1 in enumerate(obj1.getTimeInterval()): for j,t2 in enumerate(obj2.getTimeInterval()): pets[i,j] = abs(t1-t2) if useBoundingPoly: polygons1 = [pointsToShapely(obj1.getBoundingPolygon(i)) for i in obj1.getTimeInterval()] polygons2 = [pointsToShapely(obj2.getBoundingPolygon(i)) for i in obj2.getTimeInterval()] overlaps = [] for poly2 in polygons2: prep(poly2) for poly1 in polygons1: prep(poly1) overlaps.append([poly1.overlaps(poly2) for poly2 in polygons2]) smallDistances = array(overlaps) elif collisionDistanceThreshold is not None: positions1 = [p.astuple() for p in obj1.getPositions()] positions2 = [p.astuple() for p in obj2.getPositions()] distances = cdist(positions1, positions2, metric = 'euclidean') smallDistances = (distances <= collisionDistanceThreshold) else: smallDistances = array([]) if smallDistances.any(): smallPets = pets[smallDistances] petIdx = smallPets.argmin() distanceIndices = argwhere(smallDistances)[petIdx] return smallPets[petIdx], obj1.getFirstInstant()+distanceIndices[0], obj2.getFirstInstant()+distanceIndices[1] else: return None, None, None def predictPosition(self, instant, nTimeSteps, externalAcceleration = Point(0,0)): '''Predicts the position of object at instant+deltaT, at constant speed''' return predictPositionNoLimit(nTimeSteps, self.getPositionAtInstant(instant), self.getVelocityAtInstant(instant), externalAcceleration) def projectCurvilinear(self, alignments, halfWidth = 3): self.curvilinearPositions = CurvilinearTrajectory.fromTrajectoryProjection(self.getPositions(), alignments, halfWidth) def computeSmoothTrajectory(self, minCommonIntervalLength): '''Computes the trajectory as the mean of all features if a feature exists, its position is Warning work in progress TODO? not use the first/last 1-.. positions''' nFeatures = len(self.features) if nFeatures == 0: print('Empty object features\nCannot compute smooth trajectory') else: # compute the relative position vectors relativePositions = {} # relativePositions[(i,j)] is the position of j relative to i for i in range(nFeatures): for j in range(i): fi = self.features[i] fj = self.features[j] inter = fi.commonTimeInterval(fj) if inter.length() >= minCommonIntervalLength: xi = array(fi.getXCoordinates()[inter.first-fi.getFirstInstant():int(fi.length())-(fi.getLastInstant()-inter.last)]) yi = array(fi.getYCoordinates()[inter.first-fi.getFirstInstant():int(fi.length())-(fi.getLastInstant()-inter.last)]) xj = array(fj.getXCoordinates()[inter.first-fj.getFirstInstant():int(fj.length())-(fj.getLastInstant()-inter.last)]) yj = array(fj.getYCoordinates()[inter.first-fj.getFirstInstant():int(fj.length())-(fj.getLastInstant()-inter.last)]) relativePositions[(i,j)] = Point(median(xj-xi), median(yj-yi)) relativePositions[(j,i)] = -relativePositions[(i,j)] def getBoundingPolygon(self, t, shape = 'raw'): '''Returns a bounding box for the feature positions at instant bounding box format is a list of points shape is the type of output shape: - raw means the list of feature positions - rect means the bounding rectangle aligned with velocity''' if self.hasFeatures(): positions = [f.getPositionAtInstant(t) for f in self.getFeatures() if f.existsAtInstant(t)] if shape == 'rect': return Point.boundingRectangle(positions, self.getVelocityAtInstant(t)) elif shape == 'raw': return positions else: print('Unknown shape') return None else: print('Object {} has no features'.format(self.getNum())) return None def motDistanceAtInstant(self, obj, instant): '''Returns distance for computing CLEAR MOT metrics (returns an actual value, otherwise munkres does not terminate)''' return Point.distanceNorm2(self.getPositionAtInstant(instant), obj.getPositionAtInstant(instant)) ### # User Type Classification ### def classifyUserTypeSpeedMotorized(self, threshold, aggregationFunc = median, nInstantsIgnoredAtEnds = 0): '''Classifies slow and fast road users slow: non-motorized -> pedestrians fast: motorized -> cars aggregationFunc can be any function that can be applied to a vector of speeds, including percentile: aggregationFunc = lambda x: percentile(x, percentileFactor) # where percentileFactor is 85 for 85th percentile''' speeds = self.getSpeeds(nInstantsIgnoredAtEnds) if aggregationFunc(speeds) >= threshold: self.setUserType(userType2Num['car']) else: self.setUserType(userType2Num['pedestrian']) def classifyUserTypeSpeed(self, speedProbabilities, aggregationFunc = median, nInstantsIgnoredAtEnds = 0): '''Classifies road user per road user type speedProbabilities are functions return P(speed|class) in a dictionary indexed by user type names Returns probabilities for each class for simple threshold classification, simply pass non-overlapping indicator functions (membership) e.g. def indic(x): if abs(x-mu) < sigma: return 1 else: return x''' if not hasattr(self, 'aggregatedSpeed'): self.aggregatedSpeed = aggregationFunc(self.getSpeeds(nInstantsIgnoredAtEnds)) userTypeProbabilities = {} for userTypename in speedProbabilities: userTypeProbabilities[userType2Num[userTypename]] = speedProbabilities[userTypename](self.aggregatedSpeed) self.setUserType(utils.argmaxDict(userTypeProbabilities)) return userTypeProbabilities def initClassifyUserTypeHoGSVM(self, aggregationFunc, pedBikeCarSVM, bikeCarSVM = None, pedBikeSpeedThreshold = inf, bikeCarSpeedThreshold = float('Inf'), nInstantsIgnoredAtEnds = 0, homography = None, intrinsicCameraMatrix = None, distortionCoefficients = None): '''Initializes the data structures for classification TODO? compute speed for longest feature?''' self.aggregatedSpeed = aggregationFunc(self.getSpeeds(nInstantsIgnoredAtEnds)) if self.aggregatedSpeed < pedBikeSpeedThreshold or bikeCarSVM is None: self.appearanceClassifier = pedBikeCarSVM elif self.aggregatedSpeed < bikeCarSpeedThreshold: self.appearanceClassifier = bikeCarSVM else: self.appearanceClassifier = carClassifier # project feature positions if self.hasFeatures(): for f in self.getFeatures(): pp = cvutils.worldToImageProject(f.getPositions().asArray(), intrinsicCameraMatrix, distortionCoefficients, homography).tolist() f.positions = Trajectory(pp) self.userTypes = {} def classifyUserTypeHoGSVMAtInstant(self, img, instant, width, height, px, py, minNPixels, rescaleSize, orientations, pixelsPerCell, cellsPerBlock, blockNorm): '''Extracts the image box around the object (of square size max(width, height) of the box around the features, with an added px or py for width and height (around the box)) computes HOG on this cropped image (with parameters rescaleSize, orientations, pixelsPerCell, cellsPerBlock) and applies the SVM model on it''' croppedImg = cvutils.imageBox(img, self, instant, width, height, px, py, minNPixels) if croppedImg is not None and len(croppedImg) > 0: hog = cvutils.HOG(croppedImg, rescaleSize, orientations, pixelsPerCell, cellsPerBlock, blockNorm) self.userTypes[instant] = int(self.appearanceClassifier.predict(hog.reshape(1,hog.size))) else: self.userTypes[instant] = userType2Num['unknown'] def classifyUserTypeYoloAtInstant(self, instant, bboxes): '''Finds the user type based on where the feature fall at instant in detected bboxes''' userTypes = [] if self.hasFeatures(): for f in self.getFeatures(): if f.existsAtInstant(instant): p = f.getPositionAtInstant(instant) for box in bboxes: xyxy = box.xyxy[0].tolist() if p.inRectangle(xyxy[0], xyxy[2], xyxy[1], xyxy[3]): userTypes.append(coco2Types[int(box.cls.item())]) if len(userTypes) > 0: if userType2Num['cyclist'] in userTypes: self.userTypes[instant] = userType2Num['cyclist'] else: self.userTypes[instant] = utils.mostCommon(userTypes) else: self.userTypes[instant] = userType2Num['unknown'] def classifyUserTypeHoGSVM(self, minSpeedEquiprobable = -1, speedProbabilities = None, maxPercentUnknown = 0.5): '''Agregates SVM detections in each image and returns probability (proportion of instants with classification in each category) images is a dictionary of images indexed by instant With default parameters, the general (ped-bike-car) classifier will be used Considered categories are the keys of speedProbabilities''' #if not hasattr(self, 'aggregatedSpeed') or not hasattr(self, 'userTypes'): # print('Initializing the data structures for classification by HoG-SVM') # self.initClassifyUserTypeHoGSVM(aggregationFunc, pedBikeCarSVM, bikeCarSVM, pedBikeSpeedTreshold, bikeCarSpeedThreshold, nInstantsIgnoredAtEnds) # if len(self.userTypes) != self.length() and images is not None: # if classification has not been done previously # for t in self.getTimeInterval(): # if t not in self.userTypes: # self.classifyUserTypeHoGSVMAtInstant(images[t], t, homography, width, height, px, py, minNPixels, rescaleSize, orientations, pixelsPerCell, cellsPerBlock) # compute P(Speed|Class) if speedProbabilities is None or self.aggregatedSpeed < minSpeedEquiprobable: # equiprobable information from speed userTypeProbabilities = {userTypeNum: 1. for userTypeNum in speedProbabilities} else: userTypeProbabilities = {userTypeNum: speedProbabilities[userTypeNum](self.aggregatedSpeed) for userTypeNum in speedProbabilities} # compute P(Class|Appearance) nInstantsUserType = {userTypeNum: 0 for userTypeNum in userTypeProbabilities}# number of instants the object is classified as userTypename nInstantsUserType[userType2Num['unknown']] = 0 for t in self.userTypes: nInstantsUserType[self.userTypes[t]] += 1 #nInstantsUserType.get(self.userTypes[t], 0) + 1 # result is P(Class|Appearance) x P(Speed|Class) if nInstantsUserType[userType2Num['unknown']] < maxPercentUnknown*self.length(): # if not too many unknowns for userTypeNum in userTypeProbabilities: userTypeProbabilities[userTypeNum] *= nInstantsUserType[userTypeNum] # class is the user type that maximizes usertype probabilities if nInstantsUserType[userType2Num['unknown']] >= maxPercentUnknown*self.length() and (speedProbabilities is None or self.aggregatedSpeed < minSpeedEquiprobable): # if no speed information and too many unknowns self.setUserType(userType2Num['unknown']) else: # if too many unknowns here, probas are just speed probas self.setUserType(utils.argmaxDict(userTypeProbabilities)) def classifyUserTypeArea(self, areas, homography): '''Classifies the object based on its location (projected to image space) areas is a dictionary of matrix of the size of the image space for different road users possible locations, indexed by road user type names TODO: areas could be a wrapper object with a contains method that would work for polygons and images (with wrapper class) skip frames at beginning/end?''' print('not implemented/tested yet') if not hasattr(self, projectedPositions): if homography is not None: self.projectedPositions = obj.positions.homographyProject(homography) else: self.projectedPositions = obj.positions possibleUserTypes = {userType: 0 for userType in range(len(userTypenames))} for p in self.projectedPositions: for userTypename in areas: if areas[userTypename][p.x, p.y] != 0: possibleUserTypes[userType2Enum[userTypename]] += 1 # what to do: threshold for most common type? self.setUserType() return possibleUserTypes @staticmethod def collisionCourseDotProduct(movingObject1, movingObject2, instant): 'A positive result indicates that the road users are getting closer' deltap = movingObject1.getPositionAtInstant(instant)-movingObject2.getPositionAtInstant(instant) deltav = movingObject2.getVelocityAtInstant(instant)-movingObject1.getVelocityAtInstant(instant) return Point.dot(deltap, deltav) @staticmethod def collisionCourseCosine(movingObject1, movingObject2, instant): 'A positive result indicates that the road users are getting closer' return Point.cosine(movingObject1.getPositionAtInstant(instant)-movingObject2.getPositionAtInstant(instant), #deltap movingObject2.getVelocityAtInstant(instant)-movingObject1.getVelocityAtInstant(instant)) #deltav class Prototype(object): 'Class for a prototype' def __init__(self, filename, num, trajectoryType, nMatchings = None): self.filename = filename self.num = num self.trajectoryType = trajectoryType self.nMatchings = nMatchings self.movingObject = None def getFilename(self): return self.filename def getNum(self): return self.num def getTrajectoryType(self): return self.trajectoryType def getNMatchings(self): return self.nMatchings def getMovingObject(self): return self.movingObject def setMovingObject(self, o): self.movingObject = o def __str__(self): return '{} {} {}'.format(self.filename, self.num, self.trajectoryType) def __eq__(self, p2): return self.filename == p2.filename and self.num == p2.num and self.trajectoryType == p2.trajectoryType def __hash__(self): return hash((self.filename, self.num, self.trajectoryType)) ################## # Annotations ################## class BBMovingObject(MovingObject): '''Class for a moving object represented as a bounding box used for series of ground truth annotations using bounding boxes and for the output of Urban Tracker http://www.jpjodoin.com/urbantracker/ By default in image space Its center is the center of the box (generalize to other shapes?) (computed after projecting if homography available) ''' def __init__(self, topLeftPositions, bottomRightPositions, num = None, timeInterval = None, userType = userType2Num['unknown']): super(BBMovingObject, self).__init__(num, timeInterval, userType = userType) self.topLeftPositions = topLeftPositions.getPositions() self.bottomRightPositions = bottomRightPositions.getPositions() def computeCentroidTrajectory(self, homography = None): self.positions = self.topLeftPositions.add(self.bottomRightPositions).__mul__(0.5) if homography is not None: self.positions = self.positions.homographyProject(homography) def matchObjects(obj1, obj2, instant, matchingDistance): '''Indicates if obj matches obj2 with threshold matchingDistance Returns distance if below matchingDistance, matchingDistance+1 otherwise (returns an actual value, otherwise munkres does not terminate)''' d = Point.distanceNorm2(obj1.getPositionAtInstant(instant), obj2.getPositionAtInstant(instant)) if d < matchingDistance: return d else: return matchingDistance + 1 # TODO class to have different matching methods, eg with something like matchObjects def computeClearMOT(annotations, objects, matchingDistance, firstInstant, lastInstant, returnMatches = False, debug = False): '''Computes the CLEAR MOT metrics Reference: Keni, Bernardin, and Stiefelhagen Rainer. "Evaluating multiple object tracking performance: the CLEAR MOT metrics." EURASIP Journal on Image and Video Processing 2008 (2008) objects and annotations are supposed to be in the same space current implementation is BBMovingObject (bounding boxes) mathingDistance is threshold on matching between annotation and object TO: tracker output (objects) GT: ground truth (annotations) Output: returns motp, mota, mt, mme, fpt, gt mt number of missed GT.frames (sum of the number of GT not detected in each frame) mme number of mismatches fpt number of false alarm.frames (tracker objects without match in each frame) gt number of GT.frames if returnMatches is True, return as 2 new arguments the GT and TO matches matches is a dict matches[i] is the list of matches for GT/TO i the list of matches is a dict, indexed by time, for the TO/GT id matched at time t (an instant t not present in matches[i] at which GT/TO exists means a missed detection or false alarm) TODO: Should we use the distance as weights or just 1/0 if distance below matchingDistance? (add argument useDistanceForWeights = False)''' from munkres import Munkres munk = Munkres() dist = 0. # total distance between GT and TO ct = 0 # number of associations between GT and tracker output in each frame gt = 0 # number of GT.frames mt = 0 # number of missed GT.frames (sum of the number of GT not detected in each frame) fpt = 0 # number of false alarm.frames (tracker objects without match in each frame) mme = 0 # number of mismatches matches = {} # match[i] is the tracker track associated with GT i (using object references) if returnMatches: gtMatches = {a.getNum():{} for a in annotations} toMatches = {o.getNum():{} for o in objects} else: gtMatches = None toMatches = None for t in range(firstInstant, lastInstant+1): previousMatches = matches.copy() # go through currently matched GT-TO and check if they are still matched withing matchingDistance toDelete = [] for a in matches: if a.existsAtInstant(t) and matches[a].existsAtInstant(t): d = a.motDistanceAtInstant(matches[a], t) if d < matchingDistance: dist += d else: toDelete.append(a) else: toDelete.append(a) for a in toDelete: del matches[a] # match all unmatched GT-TO matchedGTs = list(matches.keys()) matchedTOs = list(matches.values()) costs = [] unmatchedGTs = [a for a in annotations if a.existsAtInstant(t) and a not in matchedGTs] unmatchedTOs = [o for o in objects if o.existsAtInstant(t) and o not in matchedTOs] nGTs = len(matchedGTs)+len(unmatchedGTs) nTOs = len(matchedTOs)+len(unmatchedTOs) if len(unmatchedTOs) > 0: for a in unmatchedGTs: costs.append([a.motDistanceAtInstant(o, t) for o in unmatchedTOs]) if len(costs) > 0: newMatches = munk.compute(costs) for k,v in newMatches: if costs[k][v] < matchingDistance: matches[unmatchedGTs[k]]=unmatchedTOs[v] dist += costs[k][v] if debug: print('{} '.format(t)+', '.join(['{} {}'.format(k.getNum(), v.getNum()) for k,v in matches.items()])) if returnMatches: for a,o in matches.items(): gtMatches[a.getNum()][t] = o.getNum() toMatches[o.getNum()][t] = a.getNum() # compute metrics elements ct += len(matches) mt += nGTs-len(matches) fpt += nTOs-len(matches) gt += nGTs # compute mismatches # for gt that do not appear in both frames, check if the corresponding to was matched to another gt in previous/next frame mismatches = [] for a in matches: if a in previousMatches: if matches[a] != previousMatches[a]: mismatches.append(a) elif matches[a] in list(previousMatches.values()): mismatches.append(matches[a]) for a in previousMatches: if a not in matches and previousMatches[a] in list(matches.values()): mismatches.append(previousMatches[a]) if debug: for mm in set(mismatches): print('{} {}'.format(type(mm), mm.getNum())) # some object mismatches may appear twice mme += len(set(mismatches)) if ct > 0: motp = dist/ct else: motp = None if gt > 0: mota = 1.-float(mt+fpt+mme)/gt else: mota = None return motp, mota, mt, mme, fpt, gt, gtMatches, toMatches def plotRoadUsers(objects, colors): '''Colors is a PlottingPropertyValues instance''' from matplotlib.pyplot import figure, axis figure() for obj in objects: obj.plot(colors.get(obj.userType)) axis('equal') if __name__ == "__main__": import doctest import unittest suite = doctest.DocFileSuite('tests/moving.txt') #suite = doctest.DocTestSuite() unittest.TextTestRunner().run(suite) #doctest.testmod() #doctest.testfile("example.txt") if shapelyAvailable: suite = doctest.DocFileSuite('tests/moving_shapely.txt') unittest.TextTestRunner().run(suite)