A Visual SLAM Algorithm with Improved Multi-Sensor Fusion Algorithm

To enhance the accuracy and stability of SLAM systems for building maps in complex environments such as factories, an enhanced scheme combining the YOLOv5s and ORB_SLAM3 systems with the ECA attention mechanism is introduced. First, a hybrid filter utilizing the Extended Kalman Filter (EKF) for prediction and pre-integration for correction is developed to improve the accuracy and stability of IMU and camera data fusion. Next, the ECA attention mechanism is integrated into the YOLOv5s network, and the improved YOLOv5s is incorporated into the ORB_SLAM3 system for target detection and dynamic feature point removal. Experimental results show that, compared to the traditional ORB_SLAM3 system, the proposed algorithm achieves up to a 91.41% reduction in root-mean-square error (RMSE) in dynamic scenes, significantly improving system accuracy and stability. Finally, the feasibility and real-time performance of the proposed method are validated through indoor map reconstruction experiments on the ROS robotics platform