The thesis addresses the problem of visual object tracking in combination with reinforcement learning methods and the usage of a learning curriculum. We implement the tracker ADNet, which iteratively picks actions to pursue objects. The tracker is trained using a policy gradient method for which we propose certain improvements, especially addressing the reward function and learning stability. The proposed curriculum is constructed from synthetic sequences gradually increasing in difficulty within two domains of synthetic objects. We experimentally show the benefits of using a curriculum on the speed and success of convergence. We compare the proposed method with learning from expert demonstrations and conclude that both methods yield similar results. The promising results from our work lead to further research in the field of curriculum learning and the use of synthetic sequences in visual object tracking.
|
