Research On Multiple Object Detection And Tracking Technology Based On Video Stream

Multi-object detection and tracking technology based on video streams is widely used in the field of intelligent traffic monitoring.How to quickly and accurately complete vehicle detection and tracking to obtain information such as road flow and driving trajectory is of great significance for applications such as traffic flow monitoring,path planning,and violation determination.Unlike pedestrian tracking tasks,the high-speed motion characteristics of vehicle targets pose new challenges for tracking algorithms.This thesis focuses on the deep learning model and post-processing in joint detection and tracking algorithms,and studies the problem of vehicle detection and tracking in road monitoring.The main content is as follows:Firstly,for the model of object detection and feature extraction,this thesis reduces the complexity and parameters of the model by using lightweight backbone network.In order to compensate for the decline in accuracy caused by lightweight backbone network,this thesis introduces adaptive activation function to improve the feature fusion module,and uses its adaptive adjustment ability for activated neurons to improve the overall network accuracy.Subsequently,Ada Face loss was introduced to act on feature extraction branches,optimizing the network’s ability to distinguish different targets and further improving the accuracy of the network.Secondly,in response to the common problems of jumping in tracking algorithms when facing special target states,this thesis proposes reducing the correlation matching dimension and improving the tracking output scheme,which can achieve more accurate and coherent vehicle tracking.In addition,for the problem of target miss detection in high magnification down sampling models,this thesis removes maximum pooling and adds non maximum suppression modules in the post-processing process to make more effective use of network output features.Finally,this thesis conducted simulation verification on the UA-DETRAC dataset to verify the algorithm performance.The experimental results showed that compared with other mainstream solutions,the model in this thesis effectively improved the running speed while maintaining similar accuracy.To meet the requirements of real-time tracking,this thesis uses edge development platform Jetson Xavier NX development board to accelerate the proposed algorithm and achieve the deployment of road monitoring modules.After testing,the running speed of the road monitoring module can meet the real-time application requirements and has good practical application value.