Visual Object Tracking Based On Correlation Filters

Visual object tracking is one of the most important and hard work in the computer vision community. It utilizes video sensors to obtain object videos, locates and tracks the object with the techniques of video analysis and understanding. Visual object tracking is the basis of many vision problems, it has widespread use in many applications, such as intelligent video surveillance, human-computer interaction, and Intelligent Transportation System(ITS), etc. In recent decades, researchers have proposed a large variety of tracking algorithms in the literature, making important contributions to the community of visual object tracking. However, object tracking in realistic scenarios is a difficult problem, there are many problems needed to be solved, for instance, partial occlusion, appearance changes, complex background, illumination changes, object irregular rapid movement, etc.Correlation filter based tracking algorithms model the appearance of objects using correlation filter. It can track complex objects through rotation, occlusion and other distractions at a fast speed. MOSSE(Minimum Output Sum of Squared Error) filter minimizes the output sum of square error, and produces a stable and adaptive correlation filter from a single frame image. It is robust to variations in illumination, scale, pose, and non-rigid deformations. MOSSE filter is trained on a few samples, thus, it contains non-critical feature of the object. In this dissertation, we focus on how to construct a sparse correlation filter from MOSSE filter, hence, we add L0-norm to MOSSE filter, resulting in a new optimization function. Solving the new optimization problem, we can obtain a more robust filter with only the critical feature of the visual object. We test our algorithm on the dataset of Benchmark(CVPR 2013) to demonstrate qualitatively and quantitatively that the proposed algorithm is effective.Intelligent Transportation System(ITS) is an important embodiment of modernization in the transport management system. Vehicle tracking is a significant component of ITS. In this thesis, we apply our sparse filter to vehicle tracking. In the experiment of vehicle tracking, we train several sparse filters, different in scale, for every vehicle appearing in the video respectively. After that, we correlate the filter over the searching window in the same size, calculate the PSR(The Peak-to-Sidelobe Ratio) for all the convolution results, and find the one with maximum PSR as the final vehicle tracking result, thus, we can achieve the task of multi-scale vehicle tracking.