VSLAM based on instance segmentation

In recent years, visual SLAM algorithms have achieved impressive results, but most visual SLAM algorithms still operate based on the assumption that the observed environment is static. When facing a moving object, it may affect the robustness of the entire system. This paper proposes a method that combines visual SLAM with deep learning-based instance segmentation. To run stably, vSLAM only needs to extract feature points on static objects. In traditional vSLAM, random sample consistency (RANSAC) is used to select the corresponding feature points. However, if the main part of the view is occupied by moving objects, many feature points become inappropriate, and RANSAC cannot perform well. According to our empirical research, the feature points of people in indoor environments usually cause errors in vSLAM. We add a separate thread based on ORB-SLAM2 and use the mask generated by instance segmentation to exclude the wrong feature points so that vSLAM can estimate the camera motion stably. In our experiment, we use the COCO dataset trained the instance segmentation network model, and the TUM RGB-D dataset is used to evaluate the method in this paper. Compared with the ORB-SLAM2 algorithm, our method can obtain higher accuracy.