Video Multi-Target Tracking Method Based On Depth Feature And Labeled Multi-Bernoulli Filter

Video multi-target tracking is an important branch in the field of computer vision.Nowadays,video target detection technology is improving day by day.It has a high recognition rate for the object category and object position information in a single picture,but there will still be missed detections and false detections.To solve this problem,video multi-target tracking can use video information to model the motion and appearance of the targets,and mark the identities with their respective labels to record the trajectory information,so as to reduce the problem of false detections and missed detections of the detector.Meanwhile,Video multitarget tracking needs to solve the problems of how to effectively identify new targets,how to re-identify targets,how to adapt to the changes of target characteristics,and how to maintain the target trajectory for a long distance.Aiming at these problems,this paper makes the following contributions under the framework of labeled multi-Bernoulli filter:1.Aiming at the problem of false detections and missed detections of 2D target detector,a detection optimized labeled multi-Bernoulli video multi-target tracking algorithm is proposed.The labeled multi-Bernoulli filter is applied to the video multi-target tracking to solve the problem of fragmented trajectories caused by the missed detections.In the traditional labeled multi-Bernoulli filter,it is necessary to establish a priori newborn target model.Due to the uncertainty of the positions and states of the newborn targets tracked by video multi-target,it is difficult to accurately set the newborn target model.Therefore,this paper proposes a measurement driven labeled multi-Bernoulli newborn model,which can effectively and timely identify the newborn targets in the scene.This newborn model lays a good foundation for improving the accuracy of continuous tracking of subsequent algorithms.Target feature extraction is introduced and applied to the target recognition module of feature pool.Experiments show that the algorithm has good robustness and accuracy in estimating the targets’ states and trajectories.2.A video multi-target tracking algorithm based on detection and feature fusion is proposed to solve the problem of low algorithm efficiency caused by step acquisition of target detection and target feature.By introducing detection and feature joint learning detector,the detector can output both targets’ detections and targets’ features after training,which alleviates the problem of low algorithm efficiency caused by asynchronous acquisition of target detection and target feature.The low confidence detection is set as the new component of labeled multi-Bernoulli.Compared with removing the low confidence detections directly,experiments show that this method can identify new targets more effectively and reduce the number of missed tracks.Adding target features into the data association part can effectively solve the problem that labeled multi-Bernoulli components with high survival probability cannot disappear in time.Compared with the new video multi-target tracking algorithm on MOT17 data set,the algorithm has good competitiveness.3.In order to solve the problem of false detections and missed detections in 3D video target tracking,a 3D multi-target tracking algorithm based on the labeled multi-Bernoulli filter is proposed.The labeled multi-Bernoulli filter is applied to 3D video multi-target tracking,and3D-IOU(3D intersection over union)is introduced to identify new targets,so as to achieve accurate recognition of 3D new targets and continuous estimation of missed tracking targets.Experiments show that the labeled multi-Bernoulli filter has a good tracking effect and development prospect in 3D video multi-target tracking.