Semantic-Aware Monocular Point-Line Visual Odometry in Dynamic Environments

This paper introduces an enhanced visual odometry (VO) system tailored for robust performance in complex dynamic environments, specifically urban road scenarios with object motion and varying lighting conditions that challenge VO accuracy. To address these challenges, we propose an indirect VO approach integrating neural network-based feature detection and matching, coupled with the extraction of line features that have strong geometric constraints. Specifically, we employ a semantic segmentation network to categorize objects into three classes: artificial static, natural static, and movable. Line features are derived from the edges of artificial static objects' segmentation masks, while features within movable object segments are disregarded. By employing these techniques, our VO system effectively leverages both geometric and semantic cues to enhance scene understanding and camera motion estimation. Comparative experiments conducted on the KITTI dataset demonstrate the superior performance of our approach in dynamic environments, exhibiting enhanced robustness and accuracy compared to conventional indirect VO methods.