| |
Student: Xuan Liu, M.Eng Student
Supervisor: Prof. Mark,Coates
Description:
Vehicle tracking is one of the most important automated techniques in traffic surveillance system; such tracking can consistently locate desired vehicles in each image of input video. This technique can help the traffic police detect, track and catch vehicles involved in traffic offences. Due to the complicated background of traffic video, vehicle tracking has several main challenges to face, such as background clutter, variety of targets, image resolution and image noise.
Motion Detection: According to the variety of vehicles and background clutter, we choose image differencing to detect the vehicles. The basic idea is to subtract the frame at time t from the next frame t+1. This difference could show any change or motion that has occurred during the interval while eliminating the stationary background. Consecutive frames have similar background since background is impossible to change significantly in such a short interval. And this also removes the non-moving objects in the video. As a result, image differencing is effective for motion detection in dynamic environment. However, image differencing also suffers from two well-known drawbacks caused by object speed and frame rate: foreground aperture and ghosting. Therefore, image differencing only detects parts of the object and also includes some parts correspond to the location of the object in the past. To overcome the drawbacks of image differencing, we use integral image to locate the detected object and perform region growing corresponding to two growing rules: (1) the number of edge row or column is decreasing; (2) the average number of edge row or column is bigger than an edge threshold. In this way, we can find the center of moving vehicles and adaptively adjust the size of detected objects.
Multitarget Tracking:With the detection results, we can start our tracking step by particle filters. Each new target will be assigned 30 particles. In the transition step, we use Gaussian random walk with 0 mean and a predicted variance . In the observation model, we use intensity model to calculate the likelihood function and update the importance weights of each particle. Then a resampling step is operated to eliminate the particles with very small weights. Finally in each step, we combine the results from new detections and propagated particles as our final targets. Thus half of particles are propagated from new detections. Another half of particles are propagated from the transition distribution.
|
Return to Particle Filtering Projects