Research On Target Tracking Method Based On Fusion Of Vision Information

Digital video image processing technology belongs to the multidisciplinary cross field,mainly involves probability theory,linear algebra,pattern recognition,digital image processing,informatics,computer science and other disciplines and it is currently one of the popular research fields.Among them,the target tracking is an important branch of digital video image processing and it is playing a pivotal role in daily life and national defense science and technology research,such as video surveillance system,license plate tracking and identification system,ballistic missile defense system,etc.But with the video scene is more and more complex,background light,target shape,as well as the changes of the interference information,put forward higher requirements on target tracking algorithms.But the existing target tracking algorithms still have some shortcomings and the tracking effect is not ideal,especially in the problems of long-term tracking and the fast moving target tracking,the loss rate of the existing target tracking algorithms is high and not easy to relock the target.In order to overcome the problems above,this paper proposes a target tracking framework which combines the human visual information and feature extraction.The framework is divided into three parts.First of all,this paper studies the Gaussian model of human vision,and the human eye saliency map is obtained from the Kinect hardware platform.Secondly,aiming at the problem that the computational efficiency is not high in the saliency extraction procedure,this paper proposes an algorithm that combines the saliency extraction and the feature extraction to improve the computational efficiency.Finally,in the target tracking,this paper proposes an improved Grab Cut image segmentation algorithm and combines the improved Mean Shift algorithm to track the target calibration area,which improves the tracking ability of anti-jamming and relocking the target.The proposed algorithm not only allows the computer to execute the various machine instructions,but also makes the computer close to the human eye vision in the mechanical procedure of image and video,so as to improve the robustness and flexibility of target tracking.In the first step,the feature points are computed by ORB(Oriented FAST and Rotated BRIEF)algorithm,and then the convex hull is computed according to the feature points.The second step is to compute the computer saliency region to obtain the computer saliency map.In the third step,the mathematical model of retina is established and the model is used to model each point in the human eye fixation point obtained by the Kinect to obtain a gaze area map closer to the human visual principle,that is,another image saliency mapthe human saliency map.The forth step,compute the three factors that affect the significant value of saliency region block for the above two kinds of saliency map separately,and then normalize them.The fifth step,fusion the computer saliency map and the human saliency map obtaining the fused saliency map.The sixth step,get the segmentation of image using improved Grab Cut to obtain the target to be tracked.The seventh step,the improved Mean Shift tracking algorithm is used to track the extracted target area.The new target tracking algorithm proposed in this paper has the following characteristics: firstly,the algorithm combines the feature extraction and target tracking and it uses the local characteristics of the image in the process of the saliency map extraction,which is more targeted and efficient.Secondly,the algorithm fuses two different saliency maps to get the final saliency map,which is better than any of the two.Finally,the algorithm for target tracking in harmony with human visual model in the process improves the anti-interference ability of tracking and relocking again,which has a further safeguard to the accuracy of target tracking.Experimental results show that for targeting area selecting automatically in initial frame in video sequences,short video sequences covering,rotating target tracking,and retargeting problem in long video sequences,the algorithm proposed in this paper has significant improvement.