Design of Self-Driving Bulldozer System

Abstract

To improve the efficiency of manual operations in the coal bunker, a completely automatic system based on multi-sensor perceptual positioning and map-based clearance planning is proposed. Firstly, the perceptual positioning link combines semantic segmentation to improve the traditional feature extraction and restricts the loop keyframe by using brute force matching, which improves the localization ability of simultaneous localization and mapping (SLAM) in the coal bunker. Compared with the incremental map maintenance strategy, a neighboring strategy is proposed to update the map in the dynamic environment of the coal bunker. Secondly, in response to the requirements of the coal bunker cleaning operation, based on the SLAM positioning results, we formulate behavior planning combined with a hierarchical finite state machine (FSM) and a grid map-based motion planning. The proposed clearing plan combines the geometric characteristics of crawler rakes, reduces the error of trajectory tracking, and improves the smoothness of trajectory curvature. The system can realize accurate real-time positioning in the bunker environment, and successfully complete the requirements of cleaning. In addition, a high simulation environment based on the unreal engine is built to verify the effectiveness and robustness of the system.