Method And Implementation Of Pedestrian Detection And Re-identification Based On Deep Learning

With the development of the country and society,public safety prevention and investigation have made people’s demand for intelligent video surveillance systems more urgent.In the era of intelligent video surveillance,artificial intelligence has become a powerful tool to strengthen social management and prevent crime.However,it is extremely difficult to perform intelligent recognition based on human face or gait in massive surveillance videos from complex scenes.Therefore,pedestrian detection and person re-identification technologies have attracted the attention of researchers.The pedestrian detection task,as its name implies,is to detect all pedestrians appearing in the image and draw the pedestrian area with a rectangular frame.Person re-identification technology is combined with pedestrian detection technology to achieve cross-camera detection and recognition,to a certain extent,to make up for the limitation of fixed camera monitoring area.Therefore,research on pedestrian detection and person re-identification technology has vital significance and value for constructing intelligent video surveillance systems.In actual surveillance scenarios,pedestrian detection’s main challenges are complex environmental backgrounds,small-sized pedestrians,or mutual occlusion between pedestrians.The main challenge of person re-identification is that the different angles of view and pose lead to significant differences between the images taken even for the same pedestrian or at the same moment.In view of the above problems,the main works of this thesis are as follow:Aiming at pedestrian detection,a pedestrian detection method with attention module SKNet is proposed.This method is based on the general target detection algorithm YOLOv3.Aiming at the situation where the YOLOv3 algorithm are prone to miss detection on occludes pedestrians and small-scale pedestrians.Through the attention mechanism module to change the weight of the convolution kernel of the original channel in the feature map,the weight of the occluded channel is decreased,and the weight of the unoccluded channel is increased,and improve detection effect of occluded pedestrian.Simultaneously,different images can get different convolution kernels,which improves the multi-scale prediction effect of YOLOv3,thereby improving the small pedestrian detection effect.In view of the problem that the model volume becomes larger and the number of parameters increases due to the attention module’s addition,model pruning enables the algorithm to improve detection speed further.In contrast,the algorithm’s accuracy is almost unchanged and meets pedestrians’real-time detection requirements in the monitoring scene.Aiming at person re-identification,a person re-identification method with combined loss function is proposed.The loss function integrates the advantages of representation learning loss and metric learning loss,and combines the improved boundary sample mining loss based on metric learning and the angular Softmax(A-Softmax)loss based on representation learning.The improved boundary sample mining loss adds a new center loss based on the boundary mining loss.The improved boundary sample mining loss reduces the intra-class distance and makes the intra-class constraint stronger.In addition,in order to solve the problem that the combined loss is difficult to converge,the BNNeck network is introduced.A system is designed to achieve the visualization of the results of the algorithm based on the pedestrian detection and person re-identification methods proposed in this thesis.The functional requirements of the system to be built are analyzed.The system’s overall scheme,including the system design architecture and working principle,is designed,and finally,based on the software platform PyQt,a system with a visual interface was built.In the set scenario,the results of the algorithm are displayed,which provides verification and interface for the application in the actual application scenario.