Fusing semantic segmentation and object detection for visual SLAM in dynamic scenes

Abstract

The assumption of static scenes limits the performance of traditional visual SLAM. Many existing solutions adopt deep learning methods or geometric constraints to solve the problem of dynamic scenes, but these schemes are either low efficiency or lack of robustness to a certain extent. In this paper, we propose a solution combining object detection and semantic segmentation to obtain the prior contours of potential dynamic objects. With this prior information, geometric constraints techniques are utilized to assist with removing dynamic feature points. Finally, the evaluation with the public datasets demonstrates that our proposed method can improve the accuracy of pose estimation and robustness of visual SLAM with no efficiency loss in high dynamic scenarios.