Pedestrian Multiple-object Tracking Based On Target Detection And Re-identification

Multi-object tracking(MOT)is an important problem in computer vision,which can be widely developed in intelligent video surveillance,automatic driving,military defense,etc.Nowadays,detection-based MOT algorithm has become the mainstream of MOT,which is developing rapidly but still facing challenges.The problems faced include illumination changes,occlusion,changes in the number of tracked targets,fast movement and high similarity of appearance,which limit the performance of the algorithm and makes it difficult to be applied in the field.With the continuous development of tracking algorithms,the detection-based MOT task gradually evolves from detection-feature extraction-based separated MOT to detection-feature extraction-based joint MOT,Therefore,the above two types of tracking methods are studied separately in this paper.The main research contents are as follows.(1)A new separated MOT network is built.Using the classical DeepSORT algorithm as the basic framework,we construct a new multi-scale target detection network by optimizing the target detection module and the feature extraction module respectively.A new feature extraction network is designed as the appearance model and a generative adversarial network is adopted to augment the training data set.The experimental results demonstrate that the target detection accuracy is improved by adding a small target detection head,and the target detection accuracy is improved by augmenting the training data set.The new appearance model trained with the augmented dataset also achieves higher TOP-1 accuracy.The improved performance of two subtasks also significantly optimizes the tracking performance metrics.(2)We analyze problems of the separated multi-objective algorithm in terms of its real time performance,and consider "real time" as a key index,and integrate detection and feature extraction branches into one network to reduce unnecessary computation by sharing features.Here,HRNet is selected as the new backbone network and the attention module is employed to enhance the features so that the extracted features are more discriminative.Finally,the loss function,which is more suitable for the re-identification task,is determined through experiments,and the tracking performance is significantly improved in accuracy and real-time performance compared to the "separated" MOT task.(3)Analyzing the joint MOT task in terms of the integration of detection and re-identification(appearance model)tasks,we find that the two subtasks contend with each other.The detection task expects the inter-class distance of the features to be large enough,while the re-identification task expects the intra-class distance of the features to be large enough.It is difficult to extract features that satisfy the expectations of both tasks,and thus the MOT task is limited.Therefore,to address the"conflict" of subtasks,a recursive cross-correlation network is constructed to decouple the shared features and learn the commonality and characteristics of the two subtasks which effectively alleviate the mutual limitation of subtask performance.The channel attention module is also designed to further enhance the features output from the re-identification branch.The performance indexes of the improved tracking model have been significantly enhanced.In summary,this paper addresses the problems of occlusion and competition among subtasks in multi-object tracking tasks,conducts research with a multi-object tracking framework of different paradigms,proposes improvement measures,and finally effectively improves the performance of multi-target tracking models.