Research On Multitarget Detection And Tracking Algorithm Based On Traffic Video

With the booming development of underlying hardware and the continuous improvement of computer performance,this provides strong motivation for the artificial intelligence algorithm based on big data fitting.These algorithms have been widely used in various fields of society,such as intelligent machine equipment,thus improving the efficiency of people’s daily lives and work.Although the intelligent transportation management system has made certain progress,there are still shortcomings such as low efficiency,slow speed,and high degree of human participation in video information acquisition.Therefore,applying more efficient,accurate,fast,and low-cost artificial intelligence algorithms in the traffic management system to achieve system intelligence has significant implications for society.Multi-target detection and tracking algorithms for video in artificial intelligence are important components in the intelligent transportation system,which can locate,identify,and track targets in traffic scenes.Although multi-target detection and tracking algorithms based on deep learning have been widely studied by the academic community,current research results still face problems such as high algorithmic costs and difficulties in identifying targets when they are mutually obscured and disappear.These problems hinder the promotion and application of related research.To address these problems,two solutions are proposed.First,a lightweight and highprecision vehicle target detection algorithm based on the YOLO framework is proposed,safeguarding the accuracy of vehicle detection and significantly reducing algorithmic costs.Second,a vehicle multi-target tracking algorithm based on the Fair Mot framework is proposed,effectively solving the problem of re-identification after target obscuration and disappearance.These two solutions can be deployed on low-cost small computer devices,efficiently processing traffic videos.For the first solution,a lightweight and high-precision vehicle target detection algorithm based on the YOLO framework is designed,using the Mobile Netv3 network as the backbone network,and improving the down-sampling and channel attention mechanisms to accurately extract target features and reduce unnecessary costs.Additionally,a feature pyramid and a single-stage headless fusion structure are designed,which can fully utilize the feature enhancement algorithm to detect multi-scale targets.Finally,SIOU is used as regression loss and Soft-NMS is used for redundant box processing,improving the algorithm’s accuracy.After design and improvement,the proposed algorithm in this thesis has achieved significant experimental results on the MS COCO dataset and the UADETRAC traffic monitoring dataset,reducing the model’s parameter and computational overheads by 64.98% and 57.14%,respectively.On the MS COCO dataset,high precision with m AP@0.5 of 58.6% is achieved,and on the UA-DETRAC traffic dataset,the accuracy of the m AP@0.5 index is improved to 70.5%,an increase of 3.52%,and the FPS is improved by 14.4%.Therefore,the first solution proposed in this thesis solves the problem of high cost and low accuracy of the algorithm.For the second solution,a multi-task detection network based on the Fair Mot framework combined with the above first solution is proposed,innovatively forming a reidentification task branch by using the deformable convolution fusion network to fuse multiscale information.The low confidence re-matching strategy of Byte Track is introduced in the tracking part to reduce algorithmic costs and improve tracking accuracy.The experimental results on the UA-DETRAC traffic monitoring dataset show that compared with the original Fair Mot-DLA34,the improved algorithm proposed in this thesis reduces the parameter overhead by 43.05%,improves FPS by 10.28%,improves MOTA by 0.57%,and improves IDs by 7.93%.Therefore,this solution further utilizes the features extracted from traffic monitoring videos by convolutional neural networks to improve the reliability and stability of the algorithm.The algorithms in this thesis are deployed on the NVIDIA AXIVAR SERIES embedded platform and have been validated through experimental applications.On the UA-DETRAC and the self-made traffic monitoring dataset,the speed of the proposed target detection algorithm reaches 15 FPS without any acceleration,and the speed of the multi-target tracking algorithm reaches 3.67fps(an increase of 55.5% over Fair Mot).This has practical significance and application value for detection tasks and vehicle flow direction determination in intelligent transportation.