Video Pedestrian Matching Across Multiple Non-overlapping Camera Views Based On Deep Learning And Their System Implementation

As information technology and intelligent city develop,video surveillance system plays an important role in many fields,which lays a great foundation for.convenient transportation and intelligent security.In our real-life scenario,however,a great number of challenges still exist.A variant of postures and different pedestrians'appearances make detection becomes a challenging problem.What's more,partly occlusion is also another significant challenge for detection in other scenarios.Traditional machine learning methods,such as HOG+SVM[20],ACF[6],DPM[8],can be used in pedestrian detection.However,hardly can they balance accuracy and efficiency greatly.Convolutional Neural Network is capable of extracting more robust pedestrian feature,which improves detection accuracy.However,ordinary Convolutional Neural Network is still unsuitable for varying illumination situation and dealing with small size target.Multiple Feature Fusion is a greater way to express target than single feature,which contains more information from target.Thus,this method,is conductive to the solution of the difficulty from varying illumination.Moreover,Multiple Feature Fusion from different dimension features obtains sufficient information in various sizes.Therefore,we propose a new deep network structure named MFRD and design a target detecting system across cameras based on a deep neural network.Besides,experiments were carried out on different datasets(Caltech[49],KITTI[50]and Campus-day-night)comparing with the other algorithms,and the experiment result shows.that the proposed method can deal well with occlusions,overlapping and improve the performance of detection.