SiamTD: Siamese twin delayed deep deterministic policy gradient for rubust Object Tracking

In the last decades, many different algorithms have been proposed to visual tracking. New trends showed that demonstrations of an agent trained by reinforcement learning can be efficiently used in visual tracking. Inspired by those achievements, we propose a novel tracking net SiameseTD(Twin delayed deep deterministic policy gradient) which combines the advantages of Siamese framework and reinforcement learning. Different from previous trackers based on Reinforcement learning, the TD algorithm we used builds on a double Q-learning net, by taking the minimum value between a pair of critics to limit over estimation problem of Q-value in the actor-critic methods. The novelty of our method is that we introduce the reinforcement learning algorithm TD into visual tracking and combine the advantages of Siamese framework. Experiments on most benchmarks show that the our tracker performs favorably better than many state-of-arts trackers.