Research On Visual Object Tracking Based On Motion Speed Awareness

Visual object tracking research on how to use computer to detect the moving track of objects in continuous image frames.Visual object tracking has important application value in video surveillance,robot vision and military reconnaissance and so on.With the deepening and perfection of the theory,the accuracy of tracking algorithm has improved significantly.However,the motion characteristics of the target object have not been fully explored and utilized.In the process of tracking,the scale of the search window and the weight of model update are often kept unchanged,and there is no adaptive adjustment according to the movement of the target,which leads to the tracking failure in the scenes of deformation,jitter and rapid movement of the target.In order to solve the above problems,this paper takes the discriminant correlation filter tracking algorithm as the main framework,studies the adaptive change of search window scale and template weight,and proposes a target tracking framework based on target motion speed awareness.The main research work of this paper is as follows:(1)To solve the problem of insufficient training samples,background awareness algorithm is introduced to expand the number of samples.The biggest difficulty of target tracking is that the training samples are only given in the first frame.When the number of samples is seriously insufficient,the trained filter is prone to over-fitting.By using the background awareness algorithm to extract the background area near the target as negative samples,the generalization performance of the filter can be enhanced.Compared with the previous method of shifting samples circulantly,the background awareness algorithm can obtain more valid samples and avoid the boundary effect in the process of model matching.(2)Aiming at the problem that trackers can’t fully extract motion characteristics of the target,the motion characteristics of the target are deeply studied,and the mathematical model of the velocity and acceleration of the target in continuous frames is proposed,and the model is further combined with the continuous convolution correlation filter.When the motion state of the target changes dramatically,the positive sample contains more interference information,and at this time,the learning rate of model update can be adjusted adaptively according to the severity of the motion state change.At the same time,the depth features with different resolutions and handcraft features are fused by cubic interpolation to improve the discrimination ability of the model.Experimental results show that the scheme proposed here can alleviate the model drift and tracking failure in scenes such as target motions abruptly,camera moves and background clutter.(3)Aiming at the problems that the search window can’t be adjusted adaptively and the target features can’t be switched according to the scene in the tracking process,a tracking algorithm based on motion speed awareness and multi-expert joint decision is proposed.Dynamically adjust the size of the search window in the next frame by the speed and acceleration of the target in the current frame to increase the probability that the target appears in the search window;At the same time,during the training process,many experts can learn the background information around the target and sample more negative samples,so as to obtain a better discrimination filter.The experimental results show that the tracker based on speed awareness and the multi-expert decision can realize robust and accurate object tracking in the case of motion abruptly and camera motion.