A Fast Visual Tracking Algorithm Based On Depth Context Model Learning

Visual tracking is to detect,extract and distinguish the moving target in video image sequences,and then to achieve the position,speed,acceleration and motion trajectory information of target,which are helpful for understanding target's motion behavior.With the increasing number of the low-price cameras and the demand of automatic video analysis,visual tracking has drawn more and more attention.However,due to more and more highly requirement for precision and speed and the increasing complexity of the tracking scenarios and target motion,designing a no-bound visual tracking algorithm still faces a big challenge,and it is an important development trend in the future.This paper is aimed at studying for single-target tracking problem.Based on the real complex scenes,this paper proposed a depth context model construction for moving target tracking to achieve good performance in robustness and accuracy.Firstly,this paper conducts a study on various theories and techniques in tracking in order to put forward an improved and optimized algorithm based on previous work.In this paper,the process of tracking is divided into several modules.Some methods which are commonly used in every module are compared and analyzed their advantages and disadvantages.The analysis procedure lays a theoretical foundation for proposing the new method.Secondly,based on the correlation filter technology,the filter template is constructed for the image sampling.The target center is located by calculating and determining the position of the maximum response.With the introducing of depth image,this paper can make up for information loss from the three dimensional space to two dimensional plane mapping.Moreover the target's surrounding region can assist to locate the target.Thirdly,utilizing the continuity and stability of depth images,this paper adopt and optimize the region growing method,which is a baseline in image segmentation.Based on the continuity and stability in depth map,the method achieves more accurate performance for target scaling.This paper proposes a method of occlusion detection and the corresponding model updating strategy based on depth images,to guarantee the long-time tracking performance effective.At last,from multiple aspects,such as comprehensive performance,scaling performance,occlusion processing performance,both qualitative and quantitative experiments demonstrate that the proposed algorithm on the accuracy and robustness performs favourably against state-of-the-art algorithms.Meanwhile,the proposed method also achieves the requirement of the real-time in speed.