Siamese global location-aware network for visual object tracking

Visual tracking is widely used in industrial systems such as vision servo systems and in intelligent robots. However, most tracking algorithms are designed without considering the balance of algorithmic efficiency and accuracy in system applications, making them less preferable for applications. This paper proposes a siamese global location-aware object tracking algorithm (SiamGLA) to address this issue. First, due to the limited performance of efficient lightweight backbone networks, this study designs an internal feature combination (IFC) module that improves feature representation with almost no additional parameters. Second, a global-aware (GA) attention module is proposed to improve the classification ability of foreground and background, which is especially important for trackers. Finally, a location-aware (LA) attention module is designed to improve the regression performance of the tracking framework. Comprehensive experiments show that SiamGLA is effective, and overcomes the drawbacks of poor robustness and weak generalization ability. When the performance reaches state-of-the-art, SiamGLA requires fewer calculations and parameters, making it more likely to be applied in practice.

Automated summary

This paper proposes a SiamGLA algorithm to address the issue of balancing algorithmic efficiency and accuracy in tracking applications. The algorithm improves feature representation through an internal feature combination (IFC) module, enhances classification ability with a global-aware (GA) attention module, and improves regression performance with a location-aware (LA) attention module. Comprehensive experiments demonstrate that SiamGLA is effective, robust and has good generalization ability, while requiring fewer calculations and parameters for practical applications.