Visual SLAM Based on Dynamic Object Detection

On the one hand, traditional visual SLAM does not consider dynamic objects in the scene, on the other hand, deep learning technology has been widely used in computer vision. This paper combines the two organically, and proposes an algorithm that uses dynamic object detection to improve the robustness of visual SLAM in a dynamic environment. Firstly, we use the object detection network integrated into the attention mechanism to detect the dynamic target in the key frame. Then, we follow the optical flow detection to further determine the dynamic feature points in the scene and eliminate them. Finally, we use the static feature points for camera tracking to achieve highly robust monocular visual SLAM. The method described in this paper can not only eliminate dynamic feature points, but also retain as many static feature points as possible. The method described in this paper is compared with the original ORB-SLAM2 algorithm and DS-SLAM algorithm, and tested with public data sets. The results show that the method described in this paper can effectively eliminate the influence of dynamic objects on the visual SLAM algorithm.