Robust Soft Actor Critic Tracking Network

Recently, various object tracking algorithms have been proposed. Reinforcement learning takes a new perspective and transforms the object tracking problem into a Markovian decision process. In this paper, we utilize the reinforcement learning algorithm SAC (Soft Actor Critic Algorithm) to train an intelligent network for object tracking. Compared with other algorithms based on AC framework, SAC utilizes the maximum entropy design to train the agent to focus more on acquiring more channels for beneficial actions, rather than being limited to acquiring a single optimal action. This design improves tracking performance while speeding up the training process. Furthermore, we replace the original feature extraction network with ResNet50 for better feature representation and demonstrate its effectiveness. We compare our algorithm with the visual tracking results of other representative trackers. Meanwhile, the excellent performance of the proposed algorithm is verified by experiments on benchmark datasets OTB100, VOT2016, VOT2018 and LaSOT.