Object Tracking in RGB and Infrared Imagery

Abstract

Object tracking is the process of locating the object throughout the frames of a video. This thesis explores tracking of an object selected by the user in RGB and infrared imagery using correlation filters. Also, we investigate illumination invariant tracking in RGB videos using median flow tracker. Additionally, we apply the correlation filter based tracker for multi object tracking to count the vehicles. The correlation filters have been widely used in computer vision for matching, detection, and tracking purposes. The basic principle of correlation filter is to learn from a set of training data to produce desired target data. The correlation filters appeal to the researchers due to its properties such as shift invariance, real-time speed, immunity to noise, and efficiency. In spite of high accuracy, the correlation filter based tracker has room for further improvements. Also, optical flow based tracker attracted tracking community recently through median flow tracking. However, there is a scope for an extension to achieve better accuracy. Thus, in this thesis, few improvements are suggested to the correlation filters for tracking applications in color and infrared imagery. The performance of a visual tracker is always degraded due to several reasons that include pose, size, appearance, illumination, occlusion, fast motion, blur, moving camera and so on. However, sudden illumination variation causes the median flow tracker to drift resulting in tracking failure. Hence, illumination invariant techniques are studied to expand the median flow tracker for robust visual tracking. This thesis considers the combination of discriminative and generative techniques by switching during uncertainty of tracked locations. The proposed technique achieves outperforming accuracy with a novel feature selection method and adaptive learning rate for correlation filter based tracker with a conditional switching to the median flow tracker. Later, the work extends combined complementary (discriminative and generative) techniques to track an object in thermal infrared imagery. Finally, the proposed techniques are tested on publicly available benchmark datasets for comparative evaluation. iiiThe thesis also presents a novel vehicle counting algorithm using an object detector combined with the correlation filter based multi object tracker. Results of the proposed algorithm are validated against the manual count.