Robust and Consistent Human Tracking Within Camera and Across Multiple Cameras

Abstract

This dissertation strives to develop a robust and consistent tracking system that tracks humans within a single camera and across multiple cameras. We present three main parts of our research: single camera tracking, multiple-camera tracking with overlapping views and multiple-camera tracking with nonoverlapping views. In single camera tracking, we focus on solving the occlusion problem in order to perform reliable tracking within a camera's view. We present an innovative method, which uses projected gradient to facilitate multiple inter-related kernels, in finding the best match during tracking under predefined constraints. The adaptive weights are applied to the kernels in order to compensate for the adverse effect introduced by occlusion. An effective scheme is also incorporated to deal with the scale change issue during the tracking. We construct the single camera tracking system by embedding the multiple-kernel tracking into a Kalman filtering-based tracking module. Several simulation results have been done to show the robustness of the proposed multiple-kernel tracking and also demonstrate that the overall system can successfully track the targets under occlusion. For multiple-camera tracking, we aim to reduce as much of the manual work as possible while providing the consistent tracking across multiple cameras. We first introduce our approach for tracking human across multiple cameras with overlapping views. The camera link model, including homography matrix, brightness transfer function and tangent transfer function, are estimated automatically given the cameras' views. Vicinity cue, color cue and edge cue are considered while matching the people across cameras. When the cameras have nonoverlapping field of views, we propose an unsupervised learning scheme to build the camera link model, including transition time distribution, brightness transfer function, region mapping matrix, region matching weights, and feature fusion weights. Our unsupervised learning scheme tolerates well the presence of outliers in the training data. The systematic integration of multiple features enables us to perform an effective re-identification across cameras. The pairwise learning and tracking manner also enhances the scalability of the system. Several experiments and comparative studies demonstrate the effectiveness of our proposed method, and the complete system has been tested in a real-world camera network scenario.