Research On Pedestrian Counting Method Based On Deep Learning In Complex Scenes

Pedestrian counting technology has become one of the branches of research in computer vision field because of its broad application prospect in security field, and the target detection and target tracking technology used in it has already represented the most research hotspot. Complexity of the application scene, including changes in lighting, pedestrian background complexity, as well as attitude changes, pedestrian occlusion and walking direction of arbitrary, these are pedestrian counting must be addressed.Although there are many pedestrian detection methods continue to put forward,but from the actual application is still a big gap.In this paper,we base on the research of pedestrian detection and target tracking technology, and clarity the deficiency of existing technology and method and emphatically studied the application of convolution neural network in pedestrian counting under practical complex scene. We propose an optimal target tracking method based on deep learning detection feedback loop, and design and implement a pedestrian counting system with high accuracy and low error. Firstly, this paper describes the research background of pedestrian counting and the research status both at home and abroad. At the same time, it introduces the target detection and tracking technology. Secondly, using the existing convolutional neural network target detection framework method, the accuracy of pedestrian head and shoulder detection and the robustness against complex scenes are improved.Finally, multi-direction and multi-area counting of pedestrian targets is realized by combining KCF (Kernel-Correlated Filter) based tracking method. Based on the above research, the pedestrian counting system via deep learning and multi-target tracking is designed in this paper.Experimental results show that the counting system has a high counting efficiency, and stable performance, high robustness, and it has a good application perspective.