Single to multiple target, multiple type visual tracking

Abstract

Visual tracking is a key task in applications such as intelligent surveillance, humancomputer interaction (HCI), human-robot interaction (HRI), augmented reality (AR), driver assistance systems, and medical applications. In this thesis, we make three main novel contributions for target tracking in video sequences. First, we develop a long-term model-free single target tracking by learning discriminative correlation filters and an online classifier that can track a target of interest in both sparse and crowded scenes. In this case, we learn two different correlation filters, translation and scale correlation filters, using different visual features. We also include a re-detection module that can re-initialize the tracker in case of tracking failures due to long-term occlusions. Second, a multiple target, multiple type filtering algorithm is developed using Random Finite Set (RFS) theory. In particular, we extend the standard Probability Hypothesis Density (PHD) filter for multiple type of targets, each with distinct detection properties, to develop multiple target, multiple type filtering, N-type PHD filter, where N ≥ 2, for handling confusions that can occur among target types at the measurements level. This method takes into account not only background false positives (clutter), but also confusions between target detections, which are in general different in character from background clutter. Then, under the assumptions of Gaussianity and linearity, we extend Gaussian mixture (GM) implementation of the standard PHD filter for the proposed N-type PHD filter termed as N-type GM-PHD filter. Third, we apply this N-type GM-PHD filter to real video sequences by integrating object detectors’ information into this filter for two scenarios. In the first scenario, a tri-GM-PHD filter is applied to real video sequences containing three types of multiple targets in the same scene, two football teams and a referee, using separate but confused detections. In the second scenario, we use a dual GM-PHD filter for tracking pedestrians and vehicles in the same scene handling their detectors’ confusions. For both cases, Munkres’s variant of the Hungarian assignment algorithm is used to associate tracked target identities between frames. We make extensive evaluations of these developed algorithms and find out that our methods outperform their corresponding state-of-the-art approaches by a large margin.