Design And Implementation Of Video-based Pedestrian Re-identification Technology

With the construction of smart city,information society and the needs of public security prevention,crime-fighting,plenty of surveillance cameras are being used in public and communities.Increasing cameras simultaneously generate massive video data,which can only be processing manually.In the face of huge video data,manual processing is costly and inefficient.To solve this problem,this paper researched the key algorithms on pedestrian detection and pedestrian re-identification,designed and implemented a personnel analysis system of key regions'entrance and exit,which can translate unstructured video data to structured data for querying.Firstly,this paper proposed a PAP pedestrian detection algorithm based on the CenterNet object detection algorithm.This algorithm followed the concept of CenterNet and regarded pedestrian detection as the problem of pedestrian's center point location.On the basis of location,two branches were added for scale regression and positioning bias regression.Some improved post-processing strategies and bounding box prediction strategies for pedestrians are also used.Experiments on the CityPersons dataset and the BUPT3000 dataset which was collected by our project team showed that the algorithm satisfied the task requirements of real-time detection.Secondly,based on PCB algorithm,this paper proposed the MV-PCB algorithm.Based on the work of PCB,this algorithm used a more powerful feature extraction network and added a branch to focus on global features from multiple perspectives.On Market1501 datasets,our algorithm get 93.4%rank-1 accuracy and 85.2 mAP,which is higher than PCB.Finally,based on the above algorithms,this paper designed and implemented a personnel analysis system of key regions'entrance and exit.The system can track the pedestrians across cameras which were under control,and it can also query the pedestrian's historical information based on the data recorded and persistently stored by the cameras.Experiments showed that the system can effectively detect person by surveillance cameras for a long time,processed video data into structured pedestrian feature data and saved it,and provided real-time pedestrian cross-camera tracking and historical information query functions.