Pedestrian Detection And Re-identification Method Research And System Implementation

In recent years,video surveillance technology is widely used in urban public places and criminal investigation tasks due to the continuous decline of video surveillance costs and the gradual improvement of monitoring results.Faced with a large amounts of video generated by the monitoring system,traditional manual surveillance methods have been unable to meet the requirements,therefore,the development of intelligent video surveillance technology is imminent,in video surveillance technology,pedestrian detection and re-identification are the focus of our research.Pedestrian detection is aim to detect the coordinate position of pedestrians who in the image or video appear,while pedestrian re-identification is designed to address the problem of identity recognition of pedestrians who captured under un-cross surveillance cameras.On the one hand,with a dramatic increase in population,pedestrian blockage often occurs in urban public places,and pedestrian occlusion problem exist in surveillance cameras;On the other hand,the pedestrian posture have various scales and vary widely,there are differences in lighting,viewpoint,and background under different cameras,pedestrian detection and re-identification don't reach the application requirements well in both accuracy and speed.Currently,deep learning is widely used in the field of computer vision,compared to the traditional methods,pedestrian detection and re-identification methods based on deep learning can overcome the above problems.Therefore,we use deep learning to design the pedestrian detector,then,combining the pedestrian detector with the pedestrian re-identification algorithm,constructing a pedestrian detection and re-identification prototype system.The specific work is as follows:1.We propose a algorithm that combining the pedestrian detector with the re-identification to address the problem of low-accuracy caused by complex background and occlusion in the real scene.The basic idea:firstly,the candidate box is extracted directly from the original image by using the YOLOv2 pedestrian detection algorithm based on convolutional neural network,predicting pedestrian position from the original image feature,transforming the pedestrian detection problem into a regression problem and achieve real-time detection of pedestrians;Next,convolutional neural network is used to extract the local characteristics of pedestrians from the image that be sent into the re-identification network and cosine softmax classifier is used for metric learning,thus,determining whether the detected pedestrian is the same person and achieve end-to-end detection and re-identification;Finally,we have optimized the training process of the model,any scale image can be input,then,fine-tune the weight of the loss function,balance the positive and negative samples,improve the detection performance.The experimental performances show that the pedestrian detection and re-identification system based on YOLOv2 has a better performance in both recognition accuracy and efficiency.2.We design and implement a prototype system that the real-time pedestrian detection and re-identification.The system consists of registration and login,image or video acquisition and analysis,setting,pedestrian detection,pedestrian re-identification and other modules,which can complete detection and re-identification tasks quickly and efficiently.The processing speed of the system achieves 23 fps,which meets the video surveillance requirements.Moreover,in the real scene testing,we found the pedestrian detection and re-identification system developed can meet the actual requirements.