Research On Multi-Object Tracking Algorithm Based On The Joint Detection And Re-Identification Neural Network

As a hot task in the field of computer vision,visual multi-object tracking(MOT)has been widely applied in various domains,such as video surveillance,traffic planning,autonomous driving,and military reconnaissance.The objective is to locate multiple objects of interest within a video and assign a unique and long-term identification to each object,thereby obtaining their motion trajectories.Multi-object tracking algorithms based on the joint detection and re-identification neural network enable endto-end training and inference of the model.However,when facing complex scenarios,such as variations in object appearance,interference from similar objects,and frequent occlusions,issues like false detections,identity switches,and trajectory discontinuities persist.Furthermore,The model size is large and the inference process is timeconsuming.This thesis aims to address the aforementioned deficiencies in the MOT algorithm’s tracking capabilities and real-time performance.Using Fair MOT algorithm as the baseline,innovation and improvement are carried out around three aspects:algorithm network framework,multi-object tracker and lightweight model,the main work is as follows:(1)An enhanced re-identification multi-object tracking algorithm based on coordinate encoding is proposed.Firstly,to satisfy the high-quality shared feature requirements for multi-object tracking,this thesis integrates channel and spatial information by incorporating a coordinate attention module in the middle layer of the feature extraction network.This makes the network more attentive to the features of interest.Secondly,to address the issue of ambiguous re-identification features extracted by the baseline algorithm,the proposed Angle-Center loss function is used to supervise the re-identification branch training,which constrains the re-identification features of objects within the same category in the angular space,close to their category feature centers,resulting in high-quality discriminative features.Finally,based on the coupling of multi-task learning,the multi-object tracking is analyzed to require low-dimensional re-identification features,and the dimension is adapted to the proposed algorithm.This balances the detection and re-identification subtasks.On the MOT17 Val dataset,the multi-object tracking metrics MOTA and IDF1 are improved by 0.7 and 1.4,respectively,and the number of IDs is reduced by 18.6% compared to the baseline algorithm.(2)A secondary data association multi-object tracking algorithm based on fusion feature similarity is proposed.Based on the improved network framework,the multiobject tracker is further optimized.On the one hand,to address the problem of insufficient feature similarity expression capability of the baseline algorithm,GR(GIo U and Re-ID)fusion feature similarity is designed,which jointly evaluates the motion feature model and re-identification feature model for a comprehensive assessment.This better reflects the similarity relationship between detection boxes and trajectory prediction boxes.On the other hand,to address the problem of poor handling of low confidence detection boxes in the previous data association process,an efficient and improved secondary data association strategy is proposed to match high and low confidence detection boxes with trajectories in two stages,which effectively avoids information loss and reduces the problem of missed detection and trajectory discontinuities caused by occlusion.On the MOT17 Val dataset,MOTA and IDF1 improved by 1.8 and 3.2 respectively compared to the baseline algorithm,and the number of IDs decreased by 32.7%.In addition,on the official MOT Challenge leaderboard,the proposed algorithm achieves the leading position in both MOT16 Test and MOT17 Test dataset tracks compared with other advanced algorithms.(3)A lightweight multi-object tracking algorithm is proposed and deployed on an embedded platform.To address the problems of high storage space occupation and high inference time consumption of the baseline algorithm,this thesis designs a lightweight feature extraction network by adaptively modifying the state-of-the-art YOLOv8 s detection network,integrating it into the joint detection and re-identification paradigm of the multi-object tracking algorithm framework.To minimize the tracking accuracy gap caused by reducing the model size,this thesis optimizes the algorithm using the components previously proposed,which maintains competitive tracking capabilities while achieving a smaller model size and faster inference.The proposed algorithm is deployed on an embedded platform with limited computing power and power.On the MOT17 Val dataset,the MOTA loss is 3.4 compared to the baseline algorithm,but the model size is reduced by 68.7% and the running frame rate is increased by 1.7 times.