DHP-SLAM: A real-time visual slam system with high positioning accuracy under dynamic environment

The traditional visual SLAM framework is assumed to be carried out in an ideal static environment. Once dynamic objects appear in the real scene, the appearance of dynamic objects greatly affects the positioning accuracy of visual SLAM system. In order to solve the above problems, a real-time multi-target tracking semantic visual SLAM system named DHP-SLAM is proposed in this paper. The dynamic visual SLAM algorithm combined with semantic instance segmentation and geometric constraint methods can eliminate the influence of high dynamic objects and potential dynamic objects, and can accurately segment objects in real time to improve the positioning of DHP-SLAM system. Secondly, multi-target tracking is integrated into the DHP-SLAM system, and the feature points of dynamic objects are eliminated by using the predicted target tracking frame when the target detection is missed, which makes the SLAM system have higher robustness and higher understanding ability of the surrounding environment. DHP-SLAM evaluated the algorithm on indoor data set TUM and outdoor data set KITTI, and conducted a large number of experiments to compare the proposed method with the state-of-the-art dynamic SLAM. In TUM indoor data set, our system has a great improvement compared with the original ORB-SLAM2 system. In KITTI data dynamic scenario, our system positioning accuracy has a great improvement in KITTI outdoor data set dynamic scenario, and it also has advantages compared with other dynamic visual SLAM systems.