Multi-target Tracking Algorithm And Application Based On Spatiotemporal Sequence Features

Multi-target tracking is a core task in the field of visual perception,which is widely used in intelligent monitoring,intelligent transportation and human-machine collaboration.In the real scene,due to the interaction between targets,severe occlusion between targets and other objects in the scene and motion blur,etc,the application of multi-target tracking technology is still extremely challenging.In this thesis,multitarget tracking algorithm is studied in human-machine collaboration scenarios,and a multi-target tracking solution that can effectively solve the problem of severe occlusion is proposed,which provides a reference for the application practice of multi-target tracking algorithm.The main contributions and innovations of this thesis can be summarized as follows:(1)In this thesis,an experimental scene consisting of four cameras installed at different angles is constructed,and a multi-target tracking dataset-HNU-T containing severe occlusion problem is constructed.The dataset sampled and synchronized the videos shot by four cameras at 25 fps to obtain 8800 real image data.The data set covers the main features of human-machine collaboration scenarios.(2)Based on the idea of spatio-temporal information fusion,this thesis proposes a multi-target tracking algorithm-STracker that integrates spatio-temporal sequence features.The time series of tracking target is fused with the features of space series to obtain the features of fusion sequence with more target discrimination.The distance matrix is calculated by fusion sequence feature and the result is obtained by bipartite graph matching with Hungarian algorithm.The tracking accuracy MOTA of STracker algorithm on HNU-T and MOT17 datasets is 85.6 and 57.2,respectively,and the trajectory quality score IDF1 is 44.4 and 50.9,respectively.(3)In order to further solve the occlusion problem,a cross-camera multi-target tracking solution based on track matching is designed combining the multi-view information of multiple cameras.The cross-camera track matching module is used to constrain the track identity under multiple cameras.The current frame detection is associated with the track segment of the total track set.This thesis innovatively proposes a global-based detection bounding box classification mechanism to improve the robustness of spatio-temporal sequence features.The tracking accuracy MOTA of the proposed cross-camera multi-target tracking solution based on track matching on HNU-T reaches 88.4,and the track quality evaluation index IDF1 reaches 61.0(16.6improvement).(4)Based on the designed cross-camera multi-target tracking algorithm,the multitarget tracking and warning system in human-machine collaboration scenarios is built.The system can accurately track the pedestrian targets in the scene under multiple cameras simultaneously,and realize the collision warning of pedestrian targets and robotic arms using the tracking results.The real-time and veracity of the whole system are analyzed and tested to prove the effectiveness of the multi-target tracking and warning system.In summary,this thesis conducts research on single-camera and cross-camera multi-target tracking algorithms in the human-machine collaboration scenario,which effectively solves the occlusion problem and improves the performance of multi-target tracking algorithms.On this basis,a multi-target tracking and early warning system under the human-machine collaboration scenario is established to explore and verify the practical application of the algorithm.