Semantic Segmentation of Point Clouds of Field Obstacle-Crossing Terrain for Multi-Legged Rescue Equipment Based on Random Forest

摘要

Multi-legged rescue equipment plays an important role in emergency rescue, military reconnaissance and military rescue due to its flexibility and adaptability. The ability of terrain recognition and scene segmentation is an important guarantee for the robot to surmount obstacles automatically, as an important part of it, point cloud semantic segmentation has also been greatly developed in recent years. However, the existing point cloud segmentation methods are all for urbanization scenes or indoor objects, and the point cloud segmentation methods for field scenes is relatively vacant. The paper aims to achieve real-time semantic segmentation for rescue equipments. First, the rule-based method is used to remove the planar terrain, and a dual-scale clustering processing framework is proposed for the remaining point clouds, which extracts the local point cloud features of small-scale clustering and fuses them into large-scale clustering, and then uses the random forest classifier to segment the scene of feature aggregation. A field point cloud data set is established, on which the experiments were carried out, in addition, compared with the decision tree, maximum likelihood and SVM classification. As a result, the random forest classification can obtain the best effect, the speed can reach 1.8s, all classes of average accuracy can reach 92.8%. The speed and accuracy are obviously better than the traditional field scene segmentation methods, which can meet the effect of real-time autonomous movement of rescue equipment, and can be used in the field real-time motion scene of large-scale engineering equipment such as rescue equipment.