Research On Pedestrian Counting System In Complex Background

The perception, analysis and understanding of scene information is the core technology of the intelligent development of information society and realizing the wisdom city. The motion pedestrian detection and tracking in the complex scene is an important part of the scenes understanding, it has wide applications in the robot autonomous navigation, natural human computer interaction and multimedia retrieval, intelligent monitoring and other fields. Pedestrian detection in moving or variable complex background has many influence factors, such as background changes, illumination and shadow et al, and the problem such as people attitude change, dimension change, angle change and occlusion.Improving the accuracy, speed and robustness of the pedestrian detection algorithm is still a challenging task now. This paper studied and realized the pedestrian counting system under the complex background from three aspects of the pedestrian detection algorithm, the multi-objective pedestrian tracking algorithm and system function design based on the analysis and summary of related research at home and abroad.According to the problem of fast pedestrian detection in complex background, this paper proposes a two cascaded fast pedestrian detection algorithm based on sparse representation. At the first stage, from the perspective of a pedestrian edge information, according to the pedestrian walk upright and symmetrical features, a vertical edge symmetrical(V_edge_sym) feature with only 2 dimensions was put forward. A large number of non-pedestrian areas can be quickly excluded by image multiscale scanning using this feature and the weak classifier based on pedestrians prior knowledge. During the second stage,the suspected pedestrian area were further accurately detected with Histograms of Oriented Gradients(HOG) features and the sparse representation classification algorithm based on the LC- KSVD dictionary learning. Experimental results show that the algorithm of this paper can ensure the detection accuracy and greatly shorten the pedestrian detection time, it also has good robustness to occlusion. In the INRIA database, the average time required for each image(640×480) was only 69 ms, the logarithmic mean missing rate is 38%, comparedwith CENTRIST+C4 algorithm and HOG+SVM algorithm missing rates relatively are decreased by 24% and 17% respectively, the detection speeds relatively are increased by 31% and 80%.For the problems of the pedestrian target shape change, drift and loss in multiple target tracking, this paper uses the method combining the pedestrian tracking and detection algorithm. Firstly the moving target areas are determined by background modeling, then the pedestrian are detected, the CEDD(Color and Edge Directivity Descriptor) characteristics of the pedestrians target are extracted, finally target similarity is judged by nearest neighbor classification,then the multi-objective pedestrian real-time tracking are realized under complex background. This algorithm can adapt to the basic morphological change of the target, such as the target size change, rotation, partly-shade or short disappearing,it can tracking target precisely and stably, which improves the robustness of the system.