Research on Robot Autonomous Inspection Visual Perception System Based on Improved YOLO V11 and Slam

Aiming at the problems existing in the robot autonomous inspection system in complex scenarios, such as visual blind spots, dynamic environment interference, and insufficient algorithm robustness, this study proposes a robot visual perception system based on improved YOLO V11 and semantic SLAM. Firstly, by introducing GhostNet lightweight convolution and the Progressive Feature Pyramid Network (P - FPN), the YOLO V11 architecture is optimized. This achieves a 30% increase in detection speed (45 FPS) on the embedded platform (Jetson AGX) while maintaining an [email protected] accuracy of 92.1%. Secondly, a semantic - enhanced SLAM framework is designed. Combining the dynamic region mask with the improved ORB - SLAM3 feature - matching strategy, the Absolute Trajectory Error (ATE) is reduced to 0.12m, and a hierarchical point - cloud map containing the semantic information of equipment status is constructed. In addition, a task - adaptive collaboration mechanism is proposed. According to the SLAM positioning information, the detection frequency and resolution mode are dynamically adjusted, reducing the system's computational resource consumption by 45%. Experiments show that on the self - built industrial inspection dataset, compared with the baseline combination (YOLOv11 + Cartographer), the small- target detection accuracy (AP - S) of this system is improved by 17.3%, and the SLAM mapping consistency score is increased by 22.8%. The actual deployment verifies the high - efficiency of the system in closed - loop tasks such as autonomous path planning, equipment status recognition, and abnormal alarm, providing an innovative solution for intelligent security in complex scenarios.