Research On Target Detection And Tracking Algorithm Based On YOLOv5 And DeepSORT

Multi-target tracking is one of the important research directions in the field of computer vision.Driven by the continuous progress of deep learning technology and social needs,multitarget tracking plays an extremely critical role in fields such as intelligent transportation,security,UAV inspection and autonomous driving.In the last decade,the performance of multitarget tracking algorithms has been improving,but accurate detection and localization of targets in realistic and complex scenarios while ensuring robustness of tracking is still a topic worthy of research.This article takes detection-based pedestrian small target tracking as the main research line,and makes targeted improvements to improve the robustness and accuracy of the algorithm based on the existing target detection and tracking algorithms,which are mainly studied as follows:(1)To address the problem that small-size targets in complex scenes have low resolution in images,are difficult to extract discriminative features,and are easily affected by environmental factors,and that existing target detection algorithms are mostly designed for regular-size targets and have poor detection effects on small-size targets,this article uses YOLOv5 m as the baseline model,based on which the network is improved in a targeted way.Firstly,a small-size target detection layer is added to improve the fusion of shallow and deep feature information and enhance the network’s ability to extract features of small-size targets.Secondly,the CA attention module is embedded in the network to help the network locate targets of interest.Finally,the original prediction head coupled on the classification and localization tasks is improved to a decoupled head on the basis of the baseline model,which accelerates the convergence of the network and improves the detection accuracy.Through experiments on the Vis Drone2019 dataset,it can be found that the improved model in this article has significantly improved in Precision,Recall and m AP,while a side-by-side comparison between the algorithm in this article and the current mainstream detection algorithms fully proves the effectiveness of the improved strategy in this article.(2)To address the problem that the Kalman filter in Deep SORT algorithm takes a uniform noise measurement scale for all target detection,which does not take into account the impact of the quality of target detection on trajectory prediction,and the data association strategy in Deep SORT algorithm lacks reasonable utilization of low confidence detection,this article uses the improved YOLOv5 m as the detector of the multi-target tracking algorithm,the Deep SORT algorithm is improved from the following aspects.First,the target motion modeling part of the original Deep SORT algorithm is improved by replacing the Kalman filter with the noise scale adaptive Kalman filter algorithm(NSA-KF)to obtain a more accurate motion state.Secondly,the simple feature extraction network in Deep SORT is replaced by the full-scale feature extraction model OSNet to extract more discriminative features to alleviate the target ID switch due to occlusion problem and to improve the robustness of target tracking at the same time.Finally,the data association part of the multi-target tracking algorithm is improved to increase the utilization of low-confidence detection frames,effectively reducing missed detections and improving the coherence of trajectories.The experiments on Vis Drone-MOT dataset show that the pedestrian small target tracking algorithm based on improved YOLOv5 m and Deep SORT in this article has been improved in terms of tracking accuracy and robustness,which reflects the effectiveness of the algorithm in this article.