Gpu accelerated algorithms for multiple object tracking
Manohar Mareboyana, Lubomı́r Řı́ha · 2012
This dissertation deals with video analysis for tracking of moving objects using a single camera. Moving object detection and object tracking are fundamental in video surveillance. The ability to track multiple objects involves developing algorithms that can track them, even during occlusion. The research reported in this dissertation focuses on two problems related to object tracking which are (i) accurate tracking of multiple objects and (ii) real-time processing rate of tracking algorithms, independent of the number of tracked objects. To improve the accuracy of object tracking, we propose a multi-level tracker consisting of a faster, but less accurate higher-level tracker that is able to reliably track objects in the absence of occlusion. However, for object-tracking during occlusion computationally more expensive but also more accurate tracking technique based on particle filterer (PF) is used. The local particle filter tracker that is proposed here exploits the objects locality and limits the particle “working area” to a small region in the image. The location and size of the region is provided by the higher level tracker. The locality awareness significantly reduces the particle drift, which is one of the most significant sources of inaccuracy of PF tracking. Since the resolution of video surveillance cameras is getting higher and higher, the amount of information that needs to be processed in every frame is also continuously increasing. To significantly improve the performance of our multilevel tracker we are proposing the GPU (Graphics Processing Unit) acceleration for all of the higher-level tracker's pixel-level operations (background modeling, foreground extraction, connected component detection), as well as for the color based particle filter (weight calculation). Today's GPUs consist of hundreds of processing units that can run in parallel and have been accepted by the High Performance Computing (HPC) community. By using the Tesla C2050 GPU accelerator, we were able to increase the number of particles processed from 200 to 3000 while maintaining a processing rate of 25 frames per second. This dissertation also presents a new, one pass, connected component detection algorithm. The algorithm is designed specifically for object tracking and therefore does not label pixels, but only extracts important features of the objects, such as location, size, etc. This algorithm was fully parallelized and can take advantage of today's multi-core CPUs as well as GPU accelerators. Our implementation is able to process VGA images in 0.5 ms, Full HD images in 1.8 ms and 256 Mpix images in 160 ms. The performance of the algorithm for very high resolution images finds applications in medical imaging.