Enhanced Motion Estimation for Autonomous Excavation: Accelerated Semantic Segmentation and ORB Features for Unstructured Environments

Abstract

An advanced method is presented for improving motion estimation in autonomous excavation operations within unstructured environments, addressing the significant challenges posed by dynamic objects and non-textured surfaces commonly encountered in earthwork scenarios. The proposed approach integrates Oriented Fast and Rotated BRIEF (ORB) feature points with semantic masks to enhance the accuracy and reliability of visual motion estimation. To obtain pixel level real-time semantic masks, a specialized semantic segmentation dataset was constructed, and a real-time segmentation method based on the DeeplabV3+ framework and MobileNetV2 backbone was implemented. The experimental platform, a modified excavator equipped with a stereo camera, was tested in an open, unstructured environment designed to simulate real-world earthmoving tasks. Extensive ablation and comparative experiments demonstrate that the proposed method substantially outperforms traditional visual Simultaneous Localization and Mapping (SLAM) algorithms particularly in scenarios involving static arm data, where the elimination of irrelevant dynamic feature points resulted in a significant improvement in both Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). Even in cases where the arm is in motion, the method maintains superior accuracy through precise segmentation of dynamic and static feature points. Additionally, the optimized CPU and GPU computation times for mask generation further contribute to a frame processing time that meets the demands of autonomous excavation tasks, underscoring its practical applicability. In summary, our solution effectively addresses motion estimation challenges in complex and dynamic environments, particularly in continuous trench excavation with dynamic objects and non-textured surfaces. This advancement minimizes human intervention and promotes the automation of excavation construction for construction machinery companies, ultimately improving project efficiency and safety.

Keywords

• Autonomous excavation
• real-time semantic segmentation
• SLAM
• visual motion estimation