Research On Video Tracking Method Based On Compressed Sensing

Video target tracking is a key technology in the field of computer vision and is the basis of all subsequent high-level visual processing,such as: context awareness,target behavior analysis and video search,etc,high-level video processing application and understanding.Due to the influence of many factors,such as: target scale changes,illumination changes,occlusion,fast movement and motion blur,etc,it is always a challenging task to seek efficient and robust target tracking model.Most of current tracking algorithms update the appearance model based on the observed samples from the neighboring frames.Although some algorithms have made good progress,they often encounter the problem,such as: loss of tracking in complex background.There are many problems to be solved in order to solve the problem of target tracking in the video,such as the problems mentioned above: scale changes,illumination changes,partial and total occlusion.There are also challenges,which is not mentioned above but faced in the process of video target tracking,such as: automatic initialization,a large number of clutters contained by foreground and background,and non-rigid target tracking problem.In fact,the latest researches have been carried out to solve the problems mentioned above.On the basis of referring and researching a large number of previous results,this paper aims at video target tracking to put forward the following two new detection algorithms on video target tracking.(1)This paper puts forward a kind of sparse cooperative target tracking algorithm.Compared with other similar algorithms,it is more adaptive,then,based on generative model to calculate the similarity between candidate samples and templates,finally,the two are integrated to calculate the confidence level of candidate samples.The online updating of the model can reduce the drift problem in the tracking and can make adaptive response to target changes.(2)This paper puts forward a kind of target tracking algorithm based on compressed sensing theory: first,randomly extract DRLBP features of images with different scales from the positive and negative sample area,then the high-dimensional feature information is projected into a low-rank compressed domain,and according to the characteristics of the compressed domain to establish appearance model.Finally,a random sparse measurement matrix is used in this paper to compress foreground and background objects.The real-time online updating is carried out in compressed domain.Ultimately,the tracking problem is transformed in binary classification problem of using Naive Bayesian Classifier.The experiment proves that the real-time online target tracking method based on compressed sensing proposed in this paper can achieve target online tracking in fast and real-time manner,while taking into account the target scale changes and occlusion,etc.(3)In terms of the above two sparse tracking algorithms,this paper carries out tracking verification respectively based on different experimental video data sets.The experiment shows that less than 1ms in average,the single frame tracking algorithm proposed in this paper is even lower,and in different tracking scene,the algorithm in this paper has strong practicability.