动态场景下基于语义和几何约束的视觉SLAM算法OA北大核心CSTPCD

Simultaneous localization and mapping(SLAM)is one of the fundamental capabilities of in-telligent mobile robots for state estimation in unknown environments.However,most visual SLAM systems rely on the assumption of a static scene,which results in severe problems of low accuracy and poor robustness in dynamic scenes.Furthermore,existing dynamic SLAM systems suffer from poor real-time performance.To address these issues,a SLAM system is proposed based on the combination of semantic and geometric constraints(DSG-SLAM),aiming at achieving real-time robust operations in dynamic scenes.DSG-SLAM integrates the GhostNet-YOLOv7 object detection network and an epipolar geometric constraint visual SLAM system in the ORB-SLAM2(Oriented FAST and Rotated BRIEF SLAM2)framework.Specifically,a parallel semantic thread is added on the basis of ORB-SLAM2 to obtain two-dimensional semantic information,and a fast dynamic feature rejection algorithm is added to the tracking thread by combining semantic and geometric constraints.Finally,the system is evaluated on the TUM public dataset and in real environments.The results show that,for high dynamic scenes,DSG-SLAM improves positioning accuracy by 94.55％compared with ORB-SLAM2,and for low dynamic scenes,the improvement is 22.99％.Furthermore,the system operates at a frequency of 30 Hz,effectively improving the positioning ac-curacy in dynamic scenes while ensuring real-time operations.