Visual Tracking Based On Correlation Filter

Object tracking is a hot research topic in the field of computer vision and image processing. For decades, a large number of scholars have been engaged in research on target tracking algorithm at home and abroad, however, due to the variability observed in the process of tracking the target information, the goal of mobility and the complexity of the background, or the background of their own shelter and other reasons, object tracking is still a very challenging problem. In recent years, correlation filter theory applied to target tracking and recognition is a hot spot, and different from traditional generation track, using correlation filter to track as a target classification problem that in the scene of the object is separated from the background, the position of maximum confidence as the target location. A major feature of the correlation filter is faster computing speed, and through the classifier and target appearance in real time update can quickly adapt to variety of the target position, pose, occlusion and similar interference, timely adjustment of the tracker state, suitable for tracking a variety of complex cases.In this paper, the fast target tracking algorithm based on correlation filtering is studied in two aspects. The main innovations are as follows:The most existing algorithms have to build the complex model and draw a large number of training samples to achieve accurate object tracking, which will produce large amount of calculation. The proposed.problem is not conducive to real-time tracking. In order to solve the problem, a real-time tracking method based on multi-channel kernel correlation filter was presented. Firstly, the target information of video frames were trained by using the nucleation ridge regression method to get the filter template. Secondly, the filter template was utilized to carry out the correlation measure for the possible area of the frame to be detected. Finally, the most relevant location was considered as the tracking result and the independent inputs of multichannel were weighted sum to solve the problem of multichannel input. A large number of comparison experiments with the existing tracking methods show that, the proposed method guarantees the tracking accuracy and its tracking speed also has obvious advantages under the different challenge factors. The proposed method avoids to extract a large number of samples by the correlation filter and use the dot product of frequency domain to instead of the correlation operation of time-domain, greatly reduces the computational complexity and makes the tracking speed completely meet the tracking demand of real-time scenario.When using a fixed size of target box only for the most relevant filter tracking algorithm, a change in the scale of the target or the target is rotated, not well adapted to the change of the target box. This paper presents a multi-feature fusion and scale adaptive correlation filter tracking algorithm, mainly correlation filter learned a set of samples containing scales and angle information to obtained classifier. The algorithm is independently calculate the target position and the target box state, the target position tracking with Algorithm 1), the target scale tracking is as follows:First, a sample of different scales at a position on a target, using the method of unconstrained correlation filter learn to obtain classifier. Then, at the target position of the current frame take a set sample with different scales and choose the best scale with a classifier. Last, update classifier. Simultaneously this paper integrate multiple features to further enhance the representation of the appearance. Through the video sequence containing varying scale experiments show that when the presence of the target scale changes, can accurately estimate the scale size of the target, and further enhance the robustness and accuracy of the algorithm.