Obstacle avoidance technology of AGV for power construction based on an improved YOLOv8 algorithm

Abstract To address obstacle avoidance safety challenges for heavy-duty automated guided vehicles (AGVs) in complex substation construction environments, this study proposes a real-time multi-sensor fusion obstacle avoidance system based on an improved YOLOv8. Traditional AGVs exhibit limitations in detecting heterogeneous obstacles such as deep pits, cable trenches, and moving personnel, including blind spots, high false alarm rates, and poor planning adaptability. By integrating a multi-scale large kernel attention module (MLKA) and a reparameterized generalized feature pyramid network (RepGFPN), the system enhances YOLOv8’s perception capabilities for multi-scale obstacles and low-texture hazardous areas. It constructs dynamic semantic maps through fusion of LiDAR point clouds and visual semantic information, while combining an improved dynamic window algorithm (DWA) with a time elastic band (TEB) planner to generate safe trajectories compliant with heavy-duty AGV motion constraints. This provides a highly robust solution for autonomous AGV operations in power construction scenarios.