Research On Video-Based Tracking In Dynamic Scenes

Video-based moving object detection (MOD) and tracking are two active subjects of computer vision. And they are also critical problems to be solved in many vision-based applications. Effective recognition and decision cannot be made without accurate object detection and successful tracking. So moving object detection and tracking are prerequisites of high level vision. Moving object tracking and tracking have played vital roles in applications such as intelligent surveillance, human-computer interaction, computer-aided driving, vehicle tracking, intelligent nursing and so on. Meanwhile, there are still many challenges while these two techniques are put to use. Camera motion and illumination variation will involve more difficulties to moving object detection. Illumination variation, heavy background clutter, nonrigid deformation and partial occlusion of tracked targets will affect the robustness of tracking. In addition, there are serious performance issues in existing multi-target tracking approaches.A new approach of moving object detection is proposed. This new approach, which is based on global motion compensation and an adaptive background model, can be used in dynamic scenes which are caused by camera motion. First, a sparse motion field is computed with a pyramidal implementation of Lucas-Kanade algorithm. Then the motion model parameters are estimated from the motion field. Previous frame was compensated with the motion model parameters. The background is then updated according to current frame and the compensated frame. Moving objects are obtained by applying background subtraction to current frame. This new approach can meet real-time requirement. It improves the accuracy of the detected objects and can avoid the stretching effect of global motion compensation plus temporal difference method.To minimize the unfavorable influences caused by illumination variation, background clutter, nonrigid deformation and partial occlusion of tracked targets, a new particle filtering tracking approach based on an improved color model is proposed. First, regularize the target region with Gaussian kernel. Second, calculate its HSV color histogram which is to be used as the target model of a particle filter. Experiments of multi-target tracking validate the approach and show its effectiveness. In addition, real-time process is achieved in two-object tracking.With the development of multi-processor and multi-core technologies, we are in an embarrassing situation: conventional sequential implemented multi-target tracking applications cannot meet real-time requirement while most of the computational capability provided by multi-processor or multi-core is wasted. To overcome the performance bottleneck, a coarse-grained parallel multi-target tracking implementation based on the OpenMP-specified shared-memory parallel programming model is explored. A list of tracked targets is maintained in a shared variable, and each target is tracked by an independent particle filter. The number of threads and the number of targets tracked by each thread are determined by the number of processing units. Compared to its corresponding optimized sequential version, the parallel implementation, which increases the number of targets in real-time tracking from 2 to 8, is of much more practical value.