Pedestrian Tracking Based On Particle Filter

Target tracking is the core technique of video content analysis in intelligent video processing.Video-based pedestrian tracking has broad application prospects in scene surveillance,human-computer interaction, behavior recognition and so on. Currently pedestrian tracking isa hot but difficult spot in the field of video object tracking, as the complexity of video sceneand the variability of pedestrian motion bring great challenges to it. Particle Filter is a targettracking algorithm following the Bayesian filtering estimation framework. It maintains a goodand stable performance even in nonlinear and non-Gaussian dynamic systems, and thereforeis widely used in video-based pedestrian tracking. This paper studies the application ofparticle filter in single-target pedestrian tracking and in multi-target pedestrian tracking forsettled scene with a fixed camera, and a multi-target tracking algorithm based on particle filteris finally presented and implemented.As target detection is the basis of target tracking, this paper first studies the detection andextraction of foreground target. Foreground mask is extracted with a mixture of Gaussianbackground model. Then shadow removing and morphological processing is applied to it.After that is the extraction of the connected components. Experiment results show that theconnected components of targets can be perfectly detected and extracted from the foregroundmask even in complex background.For single-target pedestrian tracking, the classical mean shift method and particle filtermethod are studied and tested in a complex scene. A mean shift embedded particle filter(MSEPF) algorithm is also presented and implemented. Experimental results show that theMSEPF algorithm resolves the collision of pedestrian targets much better than the mean shiftalgorithm and is qualified for single-target pedestrian tracking in complex scene.For multi-target pedestrian tracking，this paper proposes a multi-target tracking algorithmbased on particle filter. This is done in the form of blob list under the blob tracking frameworkcombined with the results of previous work. Foreground targets are carefully detected andextracted from complex background to generate a blob list for the corresponding connected components. The MSEPF tracking algorithm is adopted to update blob parameters for its goodperformance on resolving occlusion and collision. A special data association strategy ispresented to maintain and update the blob list for the random appearance, disappearance,merging and splitting of pedestrians. Experimental results show that the algorithm is a robustmulti-target pedestrian tracking algorithm in complex scene.