Research And Implementation Of Person Re-identification System Based On Pedestrian Detection

The main content of person re-identification is matching the same person under different cameras.It is widely used in intelligent surveillance,social security and criminal case detection.Person re-identification has become one of the most challenging research topics in the field of computer vision due to the changes of illumination and pedestrian posture in monitoring scenes.At present,the research of person re-identification is mostly based on the bounding box,but the actual scene image should be the complete picture and contain other pedestrians and background objects,which has more interference and is more difficult to recognize.At present,the main research idea is to divide pedestrian detection and person re-identification into two separate tasks,so the quality of pedestrian detection output pedestrian frame will directly affect the accuracy of person re-identification.Based on the above two points,instead of breaking it down into two separate tasks,we jointly handle both aspects in a single convolutional neural network,aiming at the application of person re-identification in the actual scene.In the training process,two different tasks cooperate and promote each other.The contribution of our work is as follows:(1)We analyse the basic principles of existing person re-identification algorithms,and points out the shortcomings of most current person re-identification algorithms in practical scenarios.Based on the application in real scene,this paper uses Faster R-CNN detection framework to integrate pedestrian detection into person re-identification network to achieve end-to-end learning,reduce feature reuse,and significantly improve the speed and accuracy of network computing.(2)Considering that the data in the actual monitoring scene is in the form of video,a pedestrian detection algorithm combining the inter-frame information of video sequence is proposed on the basis of the existing framework.When the current frame image is detected,the detection result of the previous frame image is put into the network,and a higher confidence is given to make the network focus on the detection area,which is more robust in solving occlusion,background interference and other problems.Then the addition of Soft-NMS(Soft Non-Maximum Suppression)improves the ability of the algorithm to solve serious occlusion problems.(3)Aiming at the problem of multi-scale matching in person re-identification after merging pedestrian detection,a person re-identification algorithm based on adaptive feature pyramid is proposed.Firstly,the validity of feature pyramids in solving multi-scale matching problem of person re-identification is verified.It is pointed out that the lowlevel features in feature pyramids damage the expression of high-level features so that the effect is not significant.By learning a set of adjustment vectors to change the distribution of low-level features,the accuracy of person re-identification in multi-scale matching is significantly improved.(4)Using Caffe+PyQt+MySQL and other tools,a person re-identification intelligent video surveillance platform is implemented.Off-line and on-line person re-identification are realized in real surveillance scenarios respectively.The test results show that the proposed algorithm has high accuracy and strong anti-jamming ability.