关键词: ORB-SLAM3, YOLO, 智能巡检, 传感器识别, 动态剔除

Abstract: Aiming at the problem of accuracy attenuation of intelligent inspection robot positioning and mapping under complex dynamic working conditions of substation, an enhanced SLAM architecture based on improved CA-YOLOv5 target detection is proposed. The multi-modal attention mechanism was used to optimize the CA-YOLOv5 network to construct a real-time dynamic target recognition framework. Through the semantic-geometric joint constraint strategy, the dynamic region mask and motion probability model are established in the feature matching stage. A dynamic feature filtering algorithm based on spatio-temporal consistency is designed to achieve the accurate elimination of dynamic interference sources and the effective retention of static scene structure in the BA optimization step. The comparison experiments between the public data set and the real dynamic scene show that the improved system reduces the positioning error by 43.7% in dynamic environment, and improves the completeness of map reconstruction by 41.5%, while maintaining good real-time processing performance. The fusion framework effectively solves the problem of mismatching and map pollution caused by dynamic elements. The scheme effectively overcomes the typical dynamic interference of substations, and provides reliable technical support for spatial perception in inspection robots, automatic driving and other fields.

Key words: ORB-SLAM3, YOLO, Intelligent inspection, Sensor recognition, Dynamic culling