Multi-Target Tracking Algorithm Based On YOLOv5 And DeepSORT In Multiple Scenarios

The target detection technology and multi-target tracking technology based on deep learning are the main research branches in the field of machine vision,and the research results in this field have a very wide range of application scenarios in the fields of intelligent driving,video surveillance,and intelligent transportation.In this thesis,we use DeepSORT multi-target tracking algorithm based on YOLOv5 detection for pedestrian multi-target tracking and vehicle multi-target tracking in two scenarios,specifically around the following aspects of research.For the field of pedestrian multi-target tracking,this thesis focuses on the problem of low multi-target tracking accuracy due to the poor robustness of existing multi-target tracking methods Re-ID in complex scenes with frequent scale changes and pedestrian occlusions.Proposes a multi-target tracking method to improve the re-identification module in YOLOv5-DeepSORT at a finer granularity level.Based on Res2 Net,build a network that continues to layer within a single residual block as the feature extraction network for the Re-ID part,which effectively improves the network’s ability to extract multi-scale features and still retains a large amount of detailed information while extracting deeper features to the target;adopt the structure of evenly divided feature maps in the output part of the backbone network to enhance the impact of local features on the overall network performance.The final reidentification model is trained on the publicly available datasets Market-1501 and Duke MTMC-re ID.The improved pedestrian tracking system is evaluated in MOT16 test sequence for tracking effectiveness.The experimental results show that compared with the default algorithm,the proposed pedestrian tracking system in this thesis improves MOTA by 5.4% and MOTP by 2.2% after training with Market-1501,and improves MOTA by 9.6% and MOTP by 2.7% after training with Duke MTMC-re ID.For the field of vehicle multi-target tracking,this thesis focuses on the problem of poor tracking accuracy due to low target detection efficiency under the influence of harsh environments such as light and weather.Based on YOLOv5,a feature extraction network with improved C3 module and a feature fusion network with fused efficient attention module are proposed to effectively improve the robustness of the vehicle detector in complex environments and fuzzy backgrounds.The improved vehicle detector is trained on the public dataset Vehicle-Dataset and the real scene dataset.The experimental results show that the proposed vehicle detection system m AP＿0.5 and m AP＿0.5:0.95 improve by 5.3% and 3.8%,respectively,and Precision and Recall improve by 6.9% and 3.2%,respectively,compared with the default algorithm.The vehicle tracker part uses Kalman filter uniform motion model to obtain the target motion information and Shuffle Net V2,a lightweight network,to obtain the target appearance information,which effectively reduces the number of network parameters and memory access,and reduces the configuration requirements of the algorithm for edge devices.Meanwhile,an efficient matching strategy is designed to enhance the robustness of the vehicle tracking system in obscured scenes,and the accurate traffic flow statistics and traffic movement direction judgment can be accomplished with the collision line counting method.