Place recognition and navigation of outdoor mobile robots based on random Forest learning with a 3D LiDAR

Bo Zhou, Yi He, Wenchao Huang, Xiang Yu, Fang Fang, Xiaomao Li

Abstract: Place recognition and loop detection play an important role in the outdoor simultaneous localization and mapping (SLAM). In this paper, we present a place recognition and navigation method based on the random forest algorithm for outdoor mobile robot with three-dimensional (3D) laser point cloud data. The 3D point cloud-based place recognition employs a global feature extractor composed of well-designed geometric and statistical features without preprocessing for the effective training and construction of the random forest classifier. Then, the environment point cloud and node map are fed into the classifier for the place recognition task. The place recognition method is subsequently applied to the loop detection of the mobile robots. To begin with, the odometry pose nodes are sorted according to their location and distance, and are then fed into the random forest classifier for loop discrimination. Eventually, loop verification based on the overlap rate of two point clouds is performed to identify the true loop. The loop detection method is combined with S4-SLAM proposed earlier by us to form the new S4-SLAM2 algorithm. Node maps constructed via S4-SLAM2 perform global re-localization in the given map by combining the place recognition method and the point cloud registration. The proposed method was verified by extensive evaluations using the KITTI dataset, as well as real-world scenarios of outdoor environments. The loop detection recall was determined as 82%, with 100% precision. The S4-SLAM2 system also exhibited high localization and mapping accuracies, with a localization output rate of 10 Hz and an average localization drift lower than 1%.