| Multi-target tracking is one of the basic tasks in computer vision.With the rapid development of deep learning in recent years,it has been widely used in intelligent video surveillance,automatic driving,human-computer interaction and other tasks.However,multi-target tracking under a single camera cannot be continuously employed to analyze the targets of interest due to factors such as visual field.Therefore,this paper physically splicing two cameras to expand the field of view,and the multi-target tracking technology under the spliced field of view is studied,so as to achieve accurate tracking of the targets for a long time.The tracking-by-detection strategy is the mainstream multi-target tracking paradigm with exceptional performance at present,and the performance of the detector will directly affect the final target tracking result.Therefore,built on the target detection algorithm,this paper develops a target detector with advanced performance to provide a highly reliable target detection frame for subsequent trackers.Then,the data association method is optimized on the basis of the existing multi-target tracking algorithm,so that it can track the target more robustly and generate the continuous target trajectory.Finally,aiming at the problem of object re-recognition in the cross-view of the spliced field of view,object matching in the overlapping field of view is carried out by the location information and appearance information of the object.The main work of this paper is as following:(1)An object detection algorithm combining background difference and SAG-YOLOv5 s is proposed.Aiming at the problem that the existing detection algorithms have poor detection effect on small targets in high-resolution images,background differential extraction of candidate regions containing potential targets under fixed cameras is introduced.The lightweight Ghost module and Sim AM three-dimensional attention mechanism is effectively fused and introduced into the YOLOv5 s network.The SAG-YOLOv5 s network with both detection accuracy and detection speed is obtained,which can be utilized to classify and accurately locate targets in candidate regions.The m AP of the algorithm on the self-built data set reaches 98.6%,which is 8.5% higher than that of the original YOLOv5 s algorithm,and the speed can still respond to the real-time requirements.(2)A single-camera multi-target tracking algorithm based on improved Byte Track is proposed.Firstly,the first stage of Byte Track will be replaced with the more advanced target detector proposed in the article.Then,based on the Byte association strategy,a lightweight target appearance model,OSNet,was introduced to assist the trajectory association of the high confidence detection box by extracting the target appearance features,and the Kalman filter state vector was modified in the process of target motion estimation,so as to obtain a higher quality prediction box.The multi-target tracking accuracy of the improved Byte Track algorithm,MOTA,reached 80.3%,which increased by3.8% compared with the original algorithm.(3)The epipolar constraint condition and the appearance features of the target are used to achieve the target matching of the overlapping view,and the system integration is carried out.When the target enters the overlapping field of view,it is necessary in order to realize cross-view re-recognition while tracking the target accurately.Therefore,in this paper,the epipolar constraint condition is used to narrow the target search scope to the corresponding epipolar lines and construct the candidate target set.Then,by sharing the appearance features extracted from the OSNet model in the tracker,the similarity of all targets in the candidate target set is calculated,so as to achieve the optimal matching of targets in the overlapping view.Finally,multi-target tracking under a single camera and target matching in an overlapping field of view are integrated systematically to realize the long-term tracking of multi-targets under spliced field of view. |
PDF Full Text Request