| Recently,with the rapid development of mobile camera technology,various mobile cameras,such as phonecamera,Go Pro,provide rich information for scene activity understanding.Compared with fixed cameras,wearable cameras have the advantage of being more flexible and convenient because they can move freely,and can track and capture targets from a better Angle in a larger range,providing more possibilities for video surveillance.However,the frequent changes of the camera’s field of vision bring great trouble to the long-term tracking of the target.Specifically,when the camera switches between different scenes,the traditional method based on spatial location and appearance matching can not meet the long-term tracking of the same target,which brings unprecedented challenges to multi-human tracking.In order to solve the above problems,this paper first proposes an Enhance Markov Decision Process to model the target tracking lifetime.Then,the state transition of the camera is simulated by using the optical flow information of the camera and the statistics of the target.Next,a frame-level classification algorithm is used to locate the start and end positions of the transition frame sequence.Finally,the image alignment algorithm is used to align the image misalignment caused by human motion before and after the transition frame sequence.By combining the above methods,we transform the proposed dynamic vision multi-human tracking problem into a joint state transition problem.To this end,we set up a professional video dataset to provide a test and evaluation benchmark for this study.In this paper,through a lot of experimental analysis,the superiority and competitiveness of this method in solving the new problem of multihuman tracking in dynamic vision are verified.It not only completes the continuous tracking of the target,but also gets a high score in the tracking accuracy.Finally,the effectiveness of the proposed method was verified by ablation experiments. |
PDF Full Text Request