Combining semantic and geometric features for robust VSLAM in dynamic environments

Visual Simultaneous Localization and Mapping (VSLAM) is the core technology for intelligent and autonomous unmanned systems such as robots and autonomous vehicle. It uses camera sensors to collect feature information in complex environments to achieve its own pose estimation, which is of great significance in the field of unmanned systems such as robots. However, the core of traditional SLAM is based on the assumption of static environment, which fundamentally contradicts the reality of dynamic scenes, leading to pose estimation errors in dynamic environments. Therefore, this paper proposes a SEM-VSLAM method that combines multitask network semantic information and geometric feature information to solve the normalization interference of dynamic objects, thereby achieving efficient and stable visual camera pose estimation. Firstly, an improved semantic segmentation network is used to assign semantic labels to each pixel in the input image, in order to better segment semantically prior objects such as pedestrians in dynamic environments. Then design an edge extraction network to obtain contour information of dynamic target objects, thereby improving the recognition accuracy of dynamic target objects. Finally, this paper proposes a motion weight estimation method that uses gaussian distribution to identify potential dynamic targets and combines maximum between-cluster variance to obtain point feature information in static regions. The comparative experimental results show that the method proposed in this paper can effectively improve the accuracy of visual camera pose estimation in the presence of dynamic object interference.