Vehicle Movement Tracking Written by Goldberg Stanislav for Vision Topics Seminar
Present Day Traffic Management Magnetic Loop Detectors Video Monitoring Systems
Magnetic Loop Detectors http://micro.magnet.fsu.edu/electromag/java/detector/index.html
Magnetic Loop Detectors Pros: Accurate counting Stable under different lighting and traffic conditions Cons: Costly : require digging up the road surface Unable to provide additional traffic parameters
Video Monitoring Systems
Video Monitoring Systems Pros: Vehicle counts and speeds Vehicle classification: { Bus, Truck, Bike, Car } Lane changes Acceleration/Deceleration Queue length for traffic jams Less costly to install then magnetic loop detectors Cons: Problems with congestion ( vehicles occlusion ) Long shadows linking vehicles together Transition between day and night
Tracking Requirements Automatic segmentation of a vehicle from a background and other vehicles so there can be a unique track associated with each vehicle Deal with variety of vehicles – motorcycles, passenger cars, buses, construction equipment, trucks, etc. Deal with a range of traffic conditions – light midday traffic, rush-hour congestion, varying speeds in different lanes. Deal with variety of lighting conditions – day, evening, night , sunny, overcast, rainy days. Real-time operation of the system
Vehicle Tracking Approaches 3D Model based Region based Active contour based Feature based
3D Model based tracking Image view is aligned with a detailed 3D model of each vehicle Pros: Easy vehicle classification { Bus, Truck, Bike ..} Cons: Memory and processing consuming approach Unrealistic to expect to be able to have detailed models for all vehicles that are found on the roadway
Region based tracking Pros: Every connected region in the image – “a blob” associated with each vehicle and then tracked over time using cross correlation. The “blobs” are found by the means of background extraction . Pros: Works well in free flowing traffic conditions Cons: Partial occlusion under congested traffic conditions leads to grouping of several vehicles into one “blob”
Active contour based tracking Representing vehicle by bounding contour of the object and dynamically update it during the tracking Pros: Reduced computational complexity compared to region based approach Cons: Partial occlusion is still a problem.
Feature based tracking Tracking not a object as a whole, but sub-features such as distinguished points or lines on the object. Pros: Partial occlusion is not a problem: some of the sub futures remains visible Cons: Additional problem to solve : which set of sub features belong to one object (grouping)
Motion Based Grouping Based on a common motion constraint aka “common fate” : sub-features that are moving rigidly are grouped together into a single vehicle. The grouping must be sensitive enough to pick up even slight acceleration or lane drift to distinguish a vehicle from the neighbors Spatial proximity (sub-features spatially close one to another ) must be taken into consideration
Motion Based Grouping: Why it’s good for vehicle tracking? For congested traffic vehicles are constantly changing their velocities to adjust to near by traffic, thus giving the grouper the information it needs to perform the segmentation. For free flowing traffic, vehicles more likely to maintain constant speed with almost no lane drift, making “common fate” grouping less useful, but there is more space between vehicles.
The Algorithm ( David Beymer et al ) Off-line camera definition Features detection Features tracking Features grouping Obtaining traffic parameters Vehicle Classification
The Algorithm
Off-line camera definitions (1) Line correspondence for the homography A projective transform H , or homography, is used to map from image coordinates (x,y) to world coordinates (X,Y)
Off-line camera definitions (1) Line correspondence for the homography H – linear transformation { rotation, scaling, shear, reflection } (x,y) rotated clockwise by theta angle and translated by (x0,y0), no scaling
Off-line camera definitions(2) Detection regions Stop Detection Area Start Detection Area
Off-line camera definitions (3) Fiducial points for camera stabilization
Features detection Corner features are chosen as sub-features Corner detector is based on image gradient
Features detection Horizontal ( x axis) differentiation can be approximated by Vertical ( y axis ) © stolen from Hagit’s Image Processing slides
Features detection
Features tracking: Kalman Filter Predictor The position and velocity of the vehicle is described by the linear state space: We assume that between the (k − 1)th and kth timestep the vehicle undergoes a constant acceleration of ak that is normally distributed, with mean 0 and standard deviation σa. From Newton laws of motion we conclude that:
Features tracking Kalman filtering predicts the area where to search for each corner feature Corner feature is found The distance between predicted and measured feature is computed If the distance is above threshold the track is rejected
Features tracking
Grouping The central principle: common motion When sub-feature is detected , it initially connected to all neighboring tracks within certain radius For all joined pairs of tracks pa(t), pb(t), relative displacement d(t)= pa(t)- pb(t) is calculated For each frame d value is computed for each edge and edge is broken if either maxtdx(t)- mintdx(t)> thresholdx maxtdy(t)- mintdy(t)> thresholdy Shadow sub-features tend to be unstable over time, so grouping eliminates them.
Grouping
Grouping : problems Oversegmentation one vehicle is grouped into several segments Overgrouping several vehicles are grouped into one segment
Real Time: hardware
Real Time: hardware PC – Intel Pentium 150 Mhz 160 MFLOPS, 260 MIPS C40 – TMS320C40, Texas Instruments Floating-Point Digital Signal Processor 40 MFLOPS, 20 MIPS Present PC - Intel Core i7-920 Quad Core Processor 63000 MFLOPS, 76000 MIPS
Results Length in min:sec G – number of reported vehicles (by algorithm ) N – number of actual vehicles (counted by human)
Results True match: A one to one match between ground truth and a group False negative: An unmatched ground truth Oversegmentation: A ground truth that is matched by more then one group False positive: An unmatched group Overgrouping: A group that matches more then one ground truth
Results Start End
Results: flow scatter plot
Results: velocity scatter plot
The end