Research Of Parallel Particle Filter Tracking Algorithm On CUDA Platform

As the core of computer vision in recent years, visual tracking has a great research value in intelligent video surveillance, robot vision and some other fields.Based on NVIDIA's CUDA platform, it can start a lot of threads working in parallel. And CUDA is widely used in image processing, video play, signal processing, artificial intelligence and some other fields.Compared with other algorithms, particle filter algorithm is more robust in some complex scenes, but it is difficult to meet the requirements of real-time tracking. So it is necessary to study the parallel particle filter algorithm and make related improvements in full use of the parallel computing advantages of GPU.Firstly, it analysis the theory of particle filter algorithm, and establishes a color model embedded particle filter algorithm, tests and compares the tracking results in different prediction methods and different color space. The result shows that the color model is robust but with a high computing complexity. Also it has some limits. Then, based on the fact that the algorithm can not satisfy the real-time requirements, it presents a GPU-based parallel particle filter algorithm to make full use of GPU parallel computing advantages. The experiments shows that it can speed up 2.5 times. Finally, as to the algorithm's limit, it introduces a more robust feature descriptor than the color model -- covariance matrix .Using covariance matrix as a feature descriptor, it implements the particle filter algorithm on GPU. Experimental results show that the algorithm is more robust.