A semantic-aware measurement-driven autonomous target search framework for UGV in complex unknown environments

摘要

Abstract Unmanned ground vehicles face numerous challenges when conducting target search in complex unknown environments, including insufficient target detection and localization capabilities, frequent repetitive paths during the search process, and low completion rates of target searches. To address these issues, we integrate target detection with autonomous exploration and propose a semantic-aware autonomous search strategy, which consists of three components: target detection and localization, local search, and global relocation. First, target detection and localization are achieved by employing YOLOv8 for object recognition and projecting the detected 2D bounding boxes onto 3D point clouds. Accurate target localization is ensured through depth histogram analysis and density-based spatial clustering of applications with noise (DBSCAN) clustering. In the following search task, we divide it into local and global stages for optimization. Second, the local search stage utilizes a semantic-aware information gain strategy within a dynamically expanded RRT framework, generating viewpoints around the robot to explore and ensuring efficient coverage of target regions in the local vicinity. Finally, in the global relocation stage, global frontiers are clustered, and a traveling salesman algorithm is applied to optimize the target visiting sequence. This ensures comprehensive target inspection on a global scale, improving search efficiency and accuracy. Comparative evaluations in various challenging simulation and real-world scenarios against state-of-the-art methods demonstrate that our approach achieves shorter search paths, reduced time, and higher target search integrity.