Scale-adaptive Visual Target Tracking Based On Kernel Correlation Filter

With the rapid development of computer vision technology,visual object tracking has become one of the important researches in computer vision.Considering the simpleness of the apparent model and the weak adaptability of the scale change of the classic algorithms during visual object tracking,relevant work is carried out based on correlation filter,and some new ideas are put forward which could be summarized as:(1)Scale adaptive visual object tracking algorithm with multi feature fusion is proposed.On the basis of the theory of visual object tracking algorithm,which is based on kernel correlation filter,the robustness of the algorithm is improved by combining the color feature with the shape feature of the target.In the part of apparent feature selection,a more complex color space named as CN is introduced as the color feature of target,which can extract the different color information of the target accurately,and has strong discrimination ability between the foreground and the background.Meanwhile,HOG feature is extracted to describe the shape of the target to cope with the geometric and optical changes during the long-term visual tracking.In the part of scale detection,in order to deal with the scale change of the target,training and detection are carried out in real-time by the theory of correlation filter with the shape feature and gray histogram of target.In addition,a multi-detectors tracking framework is proposed,which preserves and updates historical detectors,and determines the final target location based on the comprehensive decision-making ability of each detector during their work.Experimental results show that proposed method performs favorably in terms of efficiency,accuracy and robustness,and it can effectively handle complex situations such as illumination change,scale change,and etc.(2)Scale adaptive visual object tracking framework based on deep regression network is proposed.In order to cope with the scale change of the target during the long-term tracking,a reasonable network is put forward to train a general scale detection model by a large of tracking video sequences and static images.The proposed framework extracts the features of target and region of interest through two groups of the same convolution layers,and then the comparison is carried out by full connection layers to obtain the current scale information of the target.In addition,a general rule of motion smoothness during movement is put forward in this paper,and the training effect is further improved by extending the training samples reasonably under the rule support.Only one forward propagation operation is carried out with a model of off-line training during the test.Finally,a visual object tracking framework is built by combing the scale detection algorithm with the kernel correlation filter tracking algorithm.Experimental results demonstrate that proposed method performs favorably in comparison to the state-of-the-art tracking algorithms in terms of accuracy and overlap rate,and has strong robustness to the scale change and deformation of the target.