A semantic SLAM system for dynamic environments
Fangchao Hu, Qian Zong, Xurong Mou, Yin Chen, Han Wang, Miao He · 2024
With the development of robotics and self-driving vehicles, simultaneous localization and mapping technology has become critical. Visual SLAM utilizes the camera as the primary sensor and shows significant advantages in texture-rich static environments. However, conventional vision SLAM systems do not perform well in dynamic environments, such as changes in lighting, loss of camera positioning, or movement of feature points due to object motion. This study proposes a new SLAM system for dynamic environments, YOEC-DSLAM. YOEC-DSLAM combines a lightweight semantic segmentation network and a multi-view geometry approach to effectively identify and filter dynamic feature points, thereby significantly improving the accuracy and robustness of map construction. These improvements improve the real-time and accuracy of the system across the board. Experiments show that YOEC-DSLAM exhibits unique advantages in processing highly dynamic scenes.