Research on Object Tracking Technology for Orderless and Blurred Movement under Complex Scenes

University of Technology Sydney. Faculty of Engineering and Information Technology.Visual tracking is widely found in anomaly behaviour detection, self-driving, virtual reality. Recent researches reported that classic methods, including the Tracking-Learning-Detection method, the Particle Filter and the mean shift, were surpassed by deep learning in accuracy and correlation filtering in speed. However, correlation filtering can be affected by boundary effects. The conventional correlation filtering fixes the size of its detection window. When its detection window only captures partial target images due to large and sudden scale variations, the correlation filtering fails to locate the tracked target. When the target is undergoing violent shaking, motion blurs and orderless movements appear along with it. The conventional correlation filtering locks itself in the previous position of the target, and hence, the target is out of the sight of the correlation filtering. In this case, the correlation filtering drifts or fails to track. Therefore, this thesis topic is to track single-objects under complex scenes with attributes of motion blurs, orderless motions and scale variations. The main research innovation is listed as follows. (1) An approach for searching orderless movements is designed in a generative-discriminative tracking model. To address the uncertain orderless movements, a coarse-to-fine tracking framework is adopted. A spatio-temporal correlation is learned for the detection in the subsequent frames. Experiments are conducted on public databases with orderless motion attributes to validate the robustness of the proposed approach. (2) A template matching method is proposed for tracking objects with motion blurs. An effective target motion model is designed to provide supplementary appearance features. A robust similarity measure is proposed to address the outliers caused by motion blurs. Our approach outperforms other approaches in a public benchmark database with motion blurs. (3) An ensemble framework is designed to tackle scale variations. The scale of a target is estimated based on the Gaussian Particle Filtering. A high-confidence strategy is used to validate the reliability of tracking results. Our approach with hand-crafted or CNN features outperforms the methods based on correlation filtering and deep learning in databases with scale variations. To sum up, this thesis addresses boundary effects, model drifts, fixed search windows and easily interfered hand-crafted features of objects. Different trackers are proposed for tracking single-objects with orderless movements, motion blurs and scale variations. As future work, our methods can be extended to using a neural network to further improve single-object tracking models