Robust and Efficient Indoor Crowd Visual SLAM Algorithm Based on RGBD

Simultaneous Localization and Mapping is an essential element of robot position perception and environmental reconstruction. However, many existing SLAM systems rely on the static environment assumption and have limited adaptability to dynamic crowd scenarios. This article suggests a dynamic SLAM approach that Visual odometry and keyframe improve-ments, respectively to enhance the efficiency and resilience of SLAM systems in dynamic crowds environments. Firstly, a crowd environment-based object detection module is proposed, mainly detecting moving people in the scene and performing dynamic segmentation. Secondly, an adaptive keypoint criterion is designed to achieve further dynamic point detection and improve camera pose. Finally, the pose estimation error is evaluated on TUM and Bonn dynamic datasets. Results show that, compared to ORB-SLAM2, the proposed method achieves the greatest improvement of up to 98.14% in the high dynamic sequence of the TUM dataset. The improvement on Bonn dynamic dataset is at least 95.05%. This suggests the suggested approach possesses strong accuracy and robustness.