Research On Moving Object Tracking Based On Multi-camera

With the rapid development of social economy and the rapid progress of urban reform,urban traffic,public security,video monitoring and prevention in key areas have become important issues of concern,and intelligent video analysis has become a research hotspot in the industry.Target tracking is a technique to obtain the complete motion track of the target in the whole video by determining the relationship between specific targets in adjacent video sequences.Due to the limited field of view of a single camera,it is difficult to achieve the continuous tracking of a moving target,especially for many special application scenarios,the information cooperation and interaction between multiple cameras is urgent.In recent years,the video image processing technology based on deep learning has excellent performance,which helps to solve the problem of moving target tracking under the multi camera,but it still needs effective models and algorithms to solve the problems of positioning accuracy,information exchange between multi cameras and so on.Based on the deep learning theory,this paper focuses on the moving target tracking technology of the multi camera.The main research contents and innovations are as follows:(1)In order to solve the problems of single target tracking algorithm based on siamese network,such as tracking drift and inaccurate positioning,a single target tracking algorithm based on high-resolution siamese network is proposed.In order to solve this problem,high-resolution network is used to extract image features,which can fully retain the details of the image without affecting the receptive field of the model,and channel attention mechanism is integrated into the high-resolution network to highlight useful features,suppress redundant information,and further refine the extracted features.In the detection part,the idea of feature pyramid network is used to fully integrate the output features of the last three stages of the high-resolution network,and more abundant features are used to detect and track the target.In order to solve the problem of sample imbalance,the focus loss function is used as the classification loss function to make the model focus on the more difficult to distinguish positive samples and improve the performance of the model.(2)In order to get high accuracy and low miss rate of pedestrian detection results,an anti-occlusion pedestrian detection algorithm based on repulsion loss is proposed.In order to improve the accuracy of detection,the two-stage method is used as the infrastructure.In order to solve the problem of mutual occlusion between pedestrians,the repulsion loss function is used as the regression loss function to make the detection frame close to the corresponding target frame,far away from the prediction box and noncorresponding target box of adjacent targets.In order to reduce the miss rate,Soft-NMS algorithm is used to post process the output of the model and increase the condition that the detection box is suppressed.(3)In order to avoid horizontal block and complex component partition of pedestrian image and keep high recognition rate,a hierarchical bilinear module based person re-identification algorithm is proposed.The algorithm combines the advantages of classification model and verification model,and uses classification loss and verification loss to optimize the network parameters during training.Using hierarchical bilinear model to extract the local features of pedestrian image,it does not need to introduce additional model to identify the key points of the image,nor need to horizontally block the pedestrian image.The module automatically captures and recognizes the local features of pedestrian identification significance,so as to improve the accuracy of person re-identification.The experimental results show that the single target tracking algorithm based on high-resolution siamese network,the anti-occlusion pedestrian detection algorithm based on repulsion loss and the person re-identification algorithm based on hierarchical bilinear module have significant improvement in tracking accuracy,reducing the miss rate and improving the recognition rate.