SaliencyVR: Saliency Matching Based Visual Relocalization for Autonomous Vehicle

Abstract

Robust and accurate visual relocalization is a fundamental capability for autonomous vehicles and mobile robotics. Despite significant advancements in this area, there is a need to enhance the performance of most existing techniques when dealing with complex large-scale scenarios due to disruptions caused by unstable information. The presence of those unstable data not only increases unnecessary computational load but also results in erroneous matches. In response, this paper presents a novel visual relocation algorithm based on saliency matching. We first use the reprojection error and uncertainty of map points to calculate the saliency values of the image, reducing the subjective bias that exists in the artificial definition of saliency. Then, we incorporate the saliency mechanism into the local matching, effectively filtering out distracting information and generating more accurate and robust matching relationships for the final pose determination. Finally, we assess the efficacy of our proposed algorithm by comparing it to other algorithms across numerous publicly available datasets. The results confirm that our method outperforms state-of-the-art algorithms in majority of the comparison scenarios.