Research On Multiple Object Tracking Algorithm Based On Joint Detection

As a very important research topic in computer vision,multi-target tracking combines the theories and methods of target detection and data association,and is the basis for completing more advanced image understanding tasks.With the continuous update and iteration of target detection algorithms,detection-based multi-target tracking has gradually become the mainstream direction and produced different tracking modes.Among them,the One-shot mode,that is the multi-object tracking mode based on joint detection,embeds appearance features into object detection through network sharing,which better balances tracking accuracy and speed.However,this mode depends on object detection and has a high rate of missed and false detections.In addition,the detection task and the Re-ID module compete excessively,leading to frequent switching and interruption of target trajectories due to unfriendly data association methods.In response to these problems,this paper proposes the following work on detection-based multi-object tracking algorithms:(1)To address the problems of Re-ID module’s over-reliance on object detection and the resulting poor feature extraction due to detection biases,we propose a multiobject tracking algorithm that emphasizes on re-identification.Firstly,the algorithm corrects the appearance feature location extracted by the Re-ID module to reduce the introduction of other targets and background information.Secondly,we introduce a recursive gate convolutional mechanism to enhance the model’s modeling ability by capturing long-range dependencies through multi-level spatial interactions.Finally,we design an adaptive updating appearance feature data association method,which dynamically updates the target’s appearance feature to effectively reduce the number of target trajectory switches.(2)To address the issues of target trajectory switching and frequent missed detections,we propose a multi-object tracking algorithm with a multi-branch attention mechanism.Firstly,we design a lightweight multi-branch attention module to enhance features from multiple dimensions,enabling the feature extraction network to screen and extract effective feature information.Secondly,we adopt the PolyLoss loss function for the Re-ID branch to extract more accurate appearance features that differentiate different target objects.Finally,we design an adaptive feature fusion module that separates the feature fusion of the detection task and the Re-ID module task,reducing the competition between the two tasks.The adaptive feature fusion also effectively integrates the fine-grained features and high-level semantic features of the detection task and Re-ID.Experimental analysis shows that the proposed algorithm achieves high-level tracking accuracy of 61.0% and 55.5% on the MOT17 and MOT20 datasets respectively,indicating that the multi-branch attention mechanism in the multiobject tracking algorithm can effectively improve related issues and enhance tracking accuracy.(3)To address the problem of frequent trajectory switching and interruption of targets due to unfriendly data association methods,we propose a hierarchical weighted association matching algorithm.Firstly,we refine the target trajectory and current frame detection targets into hierarchical layers.Secondly,we propose a multi-feature fusion similarity matrix and adjust the weight ratio between different layer similarity matrices to generate a better similarity matrix,which improves the accuracy of association matching and reduces the number of target trajectory switches.Finally,we introduce low-score detection boxes that were previously discarded into the data association,and conduct a second association to mine the similarity between lowconfidence detection boxes and tracking trajectories,to maintain the continuity of tracking trajectories for heavily occluded targets.Experiments show that this data association method not only works effectively for our proposed algorithm,but also improves the HOTA,MOTA,and IDF1 metrics for other detection-based multi-object tracking algorithms.This article addresses the existing problems in multi-object tracking based on joint detection,including missed detections,false detections,over-reliance on detection,and data association.Relevant multi-object tracking strategies are proposed and experimentally validated.The results show that the proposed improvement strategies can effectively mitigate the impact of different problems.