Vision Multi-target Tracking Algorithm Based On 3D Motion Constraints

Visual multi-target tracking is one of the research hotspots in the field of computer vision.In the field of self driving vehicle,visual multi-target tracking is the eye of intelligent driving vehicle.In the process of driving,robust multi-target tracking algorithm can guarantee the safety of self driving vehicle.Visual multi-target tracking has the advantages of fast tracking speed,more information acquisition and low hardware cost.With the development of deep learning technology,some good appearance detectors based on convolutional neural network have been applied to the field of visual multitarget tracking.However,in self driving applications,the problem of multi-target tracking becomes more challenging due to the camera moving with autonomous vehicles,complex road environment and other factors.There are some problems in the application of the existing multi-target tracking algorithm in the automatic driving,for example,there are many kinds of targets to be tracked in the automatic driving,and the existing algorithm usually only tracks one kind of targets;in addition,in the driving of the self driving vehicle,there are a lot of cases of occlusion,missing detection and false detection of the tracked targets,and the existing algorithm is in the camera motion state which is easy to cause track discontinuity and leads to track failure.In view of this research situation,aiming at the problem of track discontinuity caused by camera movement,this paper proposes a visual multi-target tracking algorithm based on 3D motion constraint,designs a camera motion observation model and 3D spatial position prediction network to solve the problem of track discontinuity and identity drift in camera movement,and carries out experimental verification on different data sets.The implementation and performance evaluation of visual multi-target tracking algorithm are presented.The main work of this paper is as follows:(1)Read a lot of references about multi-target tracking,study the limitations of the existing visual multi-target tracking algorithm in the intelligent driving scene,use the monocular camera imaging model to estimate the camera motion,process the detection results,obtain the three-dimensional spatial position information and absolute motion information of the tracking target,reduce the interference of camera motion on the target motion observation.(2)Based on the long and short-term memory network,the motion track prediction model is designed to record the track of the target.The 3D spatial position data is used to establish the target motion model,predict the 3D coordinates of the target at the next moment,calculate the position of bounding box in the image plane,and prepare for the data association.(3)This paper proposes a multi-dimensional data association algorithm which integrates 3D spatial information and image plane information,improves the probability of correct association between detection results and tracking tracks,maintains the stable operation of tracking algorithm in camera movement,and solves the problem of tracking identity recovery after occlusion.In order to verify the effectiveness of the proposed method,I have carried out extensive verification on the open multi-target tracking data set and the video data collected by ourselves,and completed the performance evaluation of the algorithm.The experiment shows that the algorithm in this paper reduces the identity drift caused by the occlusion of the target and realizes the tracking task of multiple types and multi-target.