Research On Visual Tracking Based On Correlation Filters

Currently, as the discriminative trackers perform well in the latest Benchmark, it becomes a research hot in computer vision, in which, tracking-by-detection is widely used in video based object tracking. Among the mainstream object trackers, the correlation filters, which provide excellent tracking performance and high processing speed, gain lots of attention from the researchers. In this paper, we develop some insights into the correlation filter based object trackers.In this work, we firstly analyzed the single channel MOSSE filter, which use Gaussian function as output response and train the filters in Fourier domain, the processing speed reaches hundreds of frames per second while maintains a certain accuracy, thereby, many variants have been proposed, such as: kernel-MOSSE and CN, etc. We found the weakness after analyzing those trackers, for example, kernel-MOSSE tracker failed to work well on fast motion case since it was processed using raw pixels and lacks object representation. Based on this, we proposed a collaborative model which combines kernel based correlation filters and SOSVM(Structured Output SVM) prediction. Experimental results show that the collaborative model reduced almost 9 pixels in center location error.In this paper, we extend the linear correlation filter to multi-channel case, which can incorporate multiple feature channel descriptors into processing, so that the robustness of filter could significantly be improved. In our demonstrating system, the multi-channel HOG descriptors are utilized to represent the image patch for tracking; we found the distance precision accuracy improved 5%, in which the threshold is 20 pixels. The center location error decreased 7 pixels. In addition, in this work, we also explored the entire framework in the visual tracking system and related modules.