Efficient Multi-target Tracking In Surveillance Environment

With the social progress and the continuous development of computer technology,people have a more in-depth study of the technology in the field of computer vision.In recent years,intelligent monitoring technology is more widely used in military fields and people's lives,such as residential parking lots,traffic command and control center,and so on.Multi-target tracking technology is a hot research field in this field.It is a technique at the bottom of the target detection,which can provide useful information for the follow-up senior management through target tracking.In recent years,progress has been made in the development of multi-target tracking technology,but many problems still need to be solved,such as obstruction of targets and multi-scale issues.In this paper,multi-target tracking has been studied in view of the above problems,and a large number of experiments have been carried out after learning and researching related algorithms.The research results of this paper are as follows:(1)At present,there are many types of target tracking algorithms,which can be divided into two categories: generating method and discriminating method.In recent years,discerning class method has developed rapidly in the field of target tracking,and has a good speed and effect.In recent years,the target tracking competition has made outstanding achievements in the tracking of correlated filter(CF)class in discriminative class methods,and some high-speed target tracking algorithms have also been derived from this algorithm.This paper introduces the principle of the related filter class algorithms,and emphatically introduces the related theoretical knowledge of KCF.(2)The kernel-dependent filtering algorithm(KCF)is implemented and the algorithm is used on the MOT16 dataset to perform multi-target tracking experiments.In tracking,the multi-scale tracking target often changes in size due to the change ofcamera distance problem.Such as when the target moves from the far end of the camera to the near end or from the near end to the far end,the size of the tracking frame will not be changed accordingly,resulting in the tracking effect being poor if the tracked target is beyond or less than the range of the tracking frame the result of.Therefore,this algorithm is improved to a certain degree.Based on this algorithm,multi-scale processing is added,so that the tracking frame can continuously adapt to the target size and improve the quality of target tracking during the target tracking.(3)In the process of target tracking,occlusion between target and obstacle often occurs.However,when using KCF to do multi-target tracking,In order to solve the problem of tracking failure due to occlusion,this paper combines the target detection method of deep learning(YOLO)To detect the target to be tracked in real time and train the neural network using pedestrian and vehicle data in coco dataset and VOC dataset.When the target is tracked using the kernel-based filtering algorithm for multi-target tracking Pedestrians or vehicles can be detected while tracking the edge of the behavior,combined with a control strategy proposed in this paper to track the target re-positioning,to a large extent,to solve the tracking failure led to the problem.