Improved Particle Filter Algorithm Based On The Monte-Carlo Method

In recent years, intelligent video surveillance has been extensively studied and widely applied in many areas. Moving object detection and tracking has become a hot topic in the research of computer vision and related areas. Because most physical systems are nonlinear and non-Gaussian, particle filter algorithm (PF) has become a popular and efficient tool in tracking moving targets.In this paper the principles and characteristics of basic PF algorithm are analyzed, the key technology and main problems existed in PF applications are discussed, and the improvements on the basic PF are proposed from the following three respects.For the reason of the huge computation caused by sampling and re-sampling processes due to the big number of particles, the first improvement is made by reducing the number of particles. In the sampling procedure of every frame, it's necessary to judge whether the newly generated particle meets the tracking accuracy or not. If the new particle fits the tracking accuracy threshold, it will jump to the next frame; otherwise, it will keep generating particles until reaching the maximum number of particle sampling then re-sampling. The improved algorithm improves the efficiency, saves the computation time. Both the simulation and video tracking experiments demonstrate that the improved algorithm could track the targets successfully and achieve real-time processing.Considered that the re-sampling process plays a key role in PF computation, the second improvement is made in the re-sampling of Extended Kalman Particle Filter (EKPF) algorithm. Multinomial re-sampling is implemented instead of the importance re-sampling to ensure the variety of particles and enhance the stability of tracking. The simulation experiments show the efficiency of the improved algorithm.The Extended Kalman Filter (EKF) could track the targets good enough in non-linear system while its computation is far less than the PF's. A composite algorithm composed of the improved PF and EKF is put forward. For every frame, the EKF is first applied to track the targets and check whether the result meets the tracking accuracy or not. If the result fits the tracking accuracy threshold, it will jump to the next frame; else, the improved PF will be implemented to track the targets. This composite algorithm could get better results but cost less time. Both the simulation and video tracking experiments demonstrate the improvements of the composite algorithm..