DIP-SLAM: An RGB-D Visual Dynamic SLAM for Indoor Scenes Based on Instance Segmentation and Dynamic Probability Estimation of Feature Points

Most dynamic simultaneous localization and mapping (SLAM) systems based on object detection or semantic segmentation will directly remove feature points from detected potentially moving objects even if they are static. However, these static feature points can make the system’s localization more robust and accurate. Therefore, in order to retain more static points to make the SLAM system localization more accurate and robust, this article makes improvements from the following two aspects: the first aspect is how to accurately determine the dynamic–static attributes of feature points within semantic masks. The second aspect is how to determine the dynamic–static attributes of nonmatching points, which cannot be determined by the traditional multiview geometry method. For the first aspect, the epipolar distance distribution of feature points within the mask is first analyzed, and a linearly decreasing probability transmission model is then designed to calculate the dynamic probability of masked feature points accurately. Then, a epipolar geometry constraints-based dynamic probability transfer method (EPT) is proposed to keep more static feature points within the mask while removing dynamic feature points based on the movement of potentially moving objects. For the second aspect, a novel method combining K-means clustering, binomial logistic regression, and composite weighting method is designed to calculate the dynamic probability of nonmatching points outside the mask. Then, those nonmatching points with high dynamic probabilities can be effectively removed. Based on the above two methods, a novel dynamic visual SLAM system, termed DIP-SLAM, is presented. According to experiments conducted on the TUM RGBD public dataset, DIP-SLAM performs better than ORB-SLAM3, DynaSLAM, DS-SLAM, DSK-SLAM, SG-SLAM, and OVD-SLAM, and the average accuracy improvements in high-dynamic sequences are 92.94%, 10.49%, 46.49%, 11.93%, 23.88%, and 10.74%, respectively, while the average accuracy improvements in low-dynamic sequences are 54.02%, 31.15%, 12.18%, 29.74%, 27.51%, and 35.98%, respectively.