Stereo SLAM in Dynamic Environments Using Semantic Segmentation

As we all know, many dynamic objects appear almost continuously in the real world that are immensely capable of impairing performance majority vision-based SLAM systems based on static-world assumption. In order to improve robustness and accuracy visual high-dynamic environments, a real-time robust stereo system for scenes was proposed. To weaken influence content, moving-object detection method put forward our odometry, then semantic segmentation network combined extract pixel-level contours objects. Then, influences were significantly weakened increased markedly dynamic, complex, crowed city spaces. Following experiments with both KITTI Odometry dataset real-life scene, results showed could dramatically decrease tracking error or drift, stability high outdoor scenarios.