Research On Pedestrian Detection And Re-ID Methods Based On Deep Learning

In With the rapid development of social economy,the state and individuals pay more and more attention to personal safety and property safety.In order to improve the level of social security governance and build safe cities,security video surveillance systems have been established everywhere.In the face of the increasing number of video surveillance equipment and monitoring data,it is more and more difficult to find the specific target from the mass video.Therefore,only relying on manual screening of a large number of video information has been unable to meet the growing needs of social security,rapid and accurate intelligent video surveillance technology has become a hot research and application direction.Intelligent video surveillance system mainly filters redundant data collected by surveillance camera through image processing and pattern recognition in computer vision,and then extracts effective data for detection and analysis.Pedestrian recognition,as the core of intelligent video surveillance system,mainly solves the problem of locating the same pedestrian under different lenses.Existing pedestrian re-recognition methods are based on the matching of the cut pedestrian image,but in the real monitoring scene,there is no pedestrian marking frame,so it is necessary to detect the pedestrian first in the original video and then carry out pedestrian re-recognition.For this reason,this paper designs a pedestrian detector based on the deep learning method,and combines pedestrian detection and pedestrian recognition for research.The specific research content is as follows:(1)In the study of pedestrian detection methods,based on the excellent performance of convolutional neural network in the field of target detection,several classical network models were compared and analyzed,and finally YOLOv3 model was selected as the basic network model.In view of the poor robustness and low real-time performance of pedestrian detection method in the case of dense pedestrian and local occlusion,an automatic pedestrian detection method based on the improved YOLOv3 algorithm was proposed.The extended multi-scale prediction makes the network model more sensitive to the small and medium scale pedestrian targets.(2)In the study of person re-identification method,in order to obtain more robust pedestrian features,a SIFT and CNN feature complementary pedestrian feature representation model S-CNN was proposed.According to the characteristics of person reidentification task,a multi-task learning strategy including classification task and retrieval task is proposed.The existing classification and retrieval loss functions are explored theoretically.In the experiment,the performance of the combination of loss functions is verified,and the optimal combination term is obtained,which improves the accuracy of pedestrian recognition.(3)A method combining pedestrian detection and person re-identification is proposed.In order to adapt to the actual working environment,pedestrian detection and person re-identification are combined to complete the end-to-end person re-identification task.In a real monitoring scene,pedestrian detection results are dynamically generated into pedestrian candidate database as the input of re-identification task,and then the pedestrian re-identification network is used to match the target pedestrian,which verifies the feasibility of the algorithm.