Research On Indoor Pedestrian Multi-target Tracking Technology Based On Anchor-free

In recent years,deep learning has been instrumental in the advancement of computer vision,with numerous scholars proposing multi-target tracking algorithms based on neural networks.Multi-object tracking algorithms detect multiple objects in a video or image simultaneously and maintain their unique identification tags.Traditional two-stage multi-target tracking algorithms extract feature information of pedestrian targets in video frames using detectors before data association is completed by trackers.However,this approach requires two networks,leading to high computational requirements and low tracking efficiency.In contrast,the Center Track multi-target tracking algorithm,which adopts the anchor-free method,completes detection and tracking simultaneously in a single network,ensuring high tracking accuracy and faster reasoning speed.Nonetheless,indoor environments with narrow spaces and dense populations pose significant challenges to multi-target tracking,as numerous obstacles exist between multiple pedestrian targets,and pedestrian targets are non-rigid objects with different forms,leading to feature extraction difficulties and increased computational demands.Therefore,this thesis proposes improvements to the Center Track multi-target tracking algorithm,with a focus on enhancing tracking accuracy and efficiency through improved feature extraction networks and detection and tracking branches.Firstly,the feature extraction network DLANet is analyzed,and the attention mechanism is studied to improve feature extraction ability through the proposed SDLA multi-feature fusion network.Secondly,an improved method is proposed to enhance detection using tracking results,which enhances the detection effect of the algorithm by using tracking information as useful samples while completing data association in the tracking branch.Finally,the data set is enhanced to improve the algorithm’s generalization ability.The proposed indoor pedestrian multi-target tracking method based on the anchorfree approach was trained and tested,and compared with existing multi-target tracking methods.The model’s generalization ability was significantly improved by training with data sets processed by data enhancement techniques.The MOT17 data set was used for training and testing,including data sets in indoor environments.The experimental results demonstrate that the proposed method has higher tracking accuracy than existing methods while maintaining higher tracking efficiency,reducing the occurrence of ID switching phenomena,and improving the robustness of indoor pedestrian multi-target tracking technology.