Improved Action-Decision Network for Visual Tracking With Meta-Learning

Visual tracking is a challenging problem since it usually faces adverse factors, such as object deformation, fast motion, occlusion, and background clutter in practical applications. Reinforcement learning based Action-Decision Network (ADNet) has shown great potential for object tracking. However, ADNet has some shortcomings in optimal action selection and action reward, and suffers from inefficient tracking. To this end, an improved ADNet is proposed to enhance the tracking accuracy and efficiency. Firstly, the multi-domain training is incorporated into ADNet to further improve the feature extraction ability of its convolution layers. Then, in the reinforcement learning based training phase, both the selection criteria for optimal action and the reward function are redesigned separately to explore more appropriate action and eliminate useless action. Finally, an effective online adaptive update strategy is proposed to adapt to the appearance changes or deformation of the object during actual tracking. Specifically, meta-learning is utilized to pursue the most appropriate parameters for the network so that the parameters are closer to the optimal ones in the subsequent tracking process. Experimental results demonstrate that the proposed tracker has advantages over ADNet in terms of accuracy and efficiency.