| Tracking is a critical component of video analysis, as it provides the description of spatio-temporal relationships between observations and moving objects required by activity recognition modules. There are two tasks that we aim to address: (1) Multiple Target Tracking (MTT); (2) Tag-and-Track. The essential problem in MTT is to recover the data association between noisy observations and an unknown number of targets. To solve this problem, we proposed a Data-Driven Markov Chain Monte Carlo method to sample the data association space for a MAP (Maximum a Posterior) estimate that maximizes the spatio-temporal smoothness in both motion and appearance. Tag-and-Track is applied to track an arbitrary type of object given limited samples. The essential problem in Tag-and-Track is to establish and update an appearance model online to capture the visual signature of targets under varying circumstances, such as illumination changes, viewpoint changes and occlusions. We pose this Tag-and-Track problem as a semi-supervised learning problem, in which we aim to label a large number of unlabeled data given very limited labeled data (user selection). We propose to use two trackers combined in a Bayesian co-training framework, which unifies the CONDENSATION algorithm and co-training seamlessly. By using co-training, our method avoids learning errors reinforce themselves. In this thesis, we present the application of our method in detection and tracking of multiple moving object in airborne videos. In this application, we combine our core tracking algorithm with a set of motion detection and tracking techniques, including motion stabilization, geo-registration, etc., and demonstrate the robustness and efficiency of our methods... |
