Attentional Action-Driven Deep Networks for Visual Object Tracking

Abstract

This dissertation proposes a novel visual tracking method which is controlled by sequential actions learned by deep reinforcement learning. In the recent trackers using deep networks, tracking-by-detection scheme is adopted to select the target position with the highest matching score. The tracking-by-detection scheme achieves a good performance in a simple manner but is inefficient in exploring candidates. We propose an efficient action-driven deep tracker which is controlled by sequential actions trained by deep reinforcement learning. In contrast to the existing trackers using deep networks, the proposed tracker is designed to achieve a light computation as well as satisfactory tracking accuracy in both location and scale. The deep network to control actions is pre-trained using various training sequences and fine-tuned during tracking for online adaptation to target and background changes. The pre-training is done by utilizing deep reinforcement learning as well as supervised learning. The use of reinforcement learning enables even partially labeled data to be successfully utilized for semi-supervised learning. In addition, this dissertation tackles a tracking problem of an object interacting with other objects in a complex scene such as basketball game scenes containing various interactions among players and painting motions. For this purpose, we design a multi-agent architecture diverse interaction movements among neighboring objects near the target object. The multi-agent architecture is designed so that the main tracker could determine a proper action by utilizing the states of neighboring trackers. Through extensive evaluation, the proposed tracker is validated to achieve a competitive performance that is much faster than state-of-the-art deep network-based trackers.