RGB-D SLAM Algorithm Suitable for Dynamic Environment Based on Target Detection

When traditional SLAM algorithm is applied to position and pose estimation in dynamic scene, there are some problems such as low accuracy and poor robustness. The SLAM algorithm using semantic segmentation can achieve high accuracy in dynamic scenes, but the pixel level segmentation network used in SLAM algorithm will consume a lot of time, which can’t meet the real-time operation. In order to solve these two problems, we propose a SLAM algorithm which can run in real time in dynamic environment based on ORB-SLAM2 algorithm. Firstly, in the Tracking thread, we use the YOLO network to detect the dynamic objects, and use the double threshold method to divide the feature points in the detection frame into static feature points and dynamic feature points. In order to ensure the real-time performance, we parallel the target detection and feature point extraction. Images with dynamic feature points larger than static ones cannot be keyframes. In the Local Mapping thread, the motion consistency of the generated map points is checked to further screen the dynamic feature points. The results show that our algorithm can achieve state-of-the-art accuracy in dynamic environment, and the average processing time of a frame is 50ms.