Research On Multi-Object Visual Tracking Method Based On Optimized DeepSORT

With the development of deep learning and the improvement of computing power,deep learning-based multi-object visual tracking algorithms have developed rapidly and are widely used in pedestrian detection and tracking fields,with high academic research and practical value.This study focuses on the problem of pedestrian multi-object tracking in complex scenes,proposes an improved YOLOX and improved DeepSORT combined multi-object tracking algorithm based on object detection theory and object tracking theory.The specific research content is as follows:(1)An improved object detection algorithm based on YOLOX is proposed.First,to address the problem that the object detection network performs poorly in complex backgrounds and occlusions,this study embeds the CA attention mechanism into the YOLOX backbone network after comparing several attention mechanisms,which enhances the feature extraction ability of the network and improves pedestrian object detection accuracy in complex environments;Second,to address the problems of imbalanced positive and negative samples and slow network convergence,this study replaces the original YOLOX’s cross-entropy loss function with the Focal Loss loss function to strengthen the network’s emphasis on difficult-to-classify samples.Additionally,the DIOU loss function is used to replace the original YOLOX IOU loss function,improving the localization accuracy and convergence speed of the network.Finally,to address the issue of poor small object detection performance of the network,this study adds a large-scale prediction layer based on the original YOLOX’s three-scale prediction layer,thereby outputting more detailed spatial and position information and improving the network’s ability to detect small objects.(2)An improved DeepSORT algorithm is proposed.By incorporating the pedestrian re-identification network WRN,DeepSORT effectively enhances tracking performance and reduces the frequency of identity number changes.However,WRN has problems such as large parameter size and low performance.To address these issues,this study designs a new pedestrian re-identification network,WRN-Ghost,based on the main idea and feature pyramid FPN of Ghost Net,and trains WRN-Ghost using the pedestrian re-identification dataset Market1501.Finally,the trained WRN-Ghost model is used to replace the original DeepSORT’s pedestrian re-identification network,effectively enhancing the ability of DeepSORT algorithm to extract pedestrian appearance features and improving its accuracy.(3)Finally,an improved pedestrian multi-object tracking algorithm is obtained by optimizing the detection network of DeepSORT with the improved YOLOX algorithm.Then,through ablative experiments,the performance of the proposed algorithm and the effectiveness of the improvement strategy are verified.The experimental results show that compared with the original algorithm,the pedestrian multi-object tracking algorithm designed in this study has significantly improved detection accuracy and speed,and can better handle problems such as identity number changes due to occlusion,even in complex environments,it can maintain good tracking performance.