Siamese Network Tracker by Attention Module and Relation Detector Module

Abstract

In recent years, object tracking techniques based on Siamese networks have shown excellent tracking performance. However, in the tracking process, there will be many similar objects, and it is easy to track the wrong object due to the weak discriminative ability of the network. At the same time, the classification and regression of SiamRPN ++ are usually optimized independently, which will cause a mismatch problem, that is, the location with the highest classification confidence is not necessarily the object. To address these problems, we proposed a Siamese network tracker by attention module and relation detector module (SiamAR). First, we introduce a multi-scale attention mechanism in SiamRPN++ to capture information at different scales, and fuse spatial attention and channel attention to improving the ability to learn feature information. Not only different receptive fields are obtained, but also useful features are selectively focused and less useful features are suppressed. In order not to affect the computational efficiency, the method of grouping parallel computing is used. Secondly, we add a relation detector module to our tracker to filter out distractors from the background and distinguish the object in the cluttered background. Experiment results show that our algorithm out-performs several well-known tracking algorithms in terms of tracking accuracy and robustness.