LK-ORB-SLAM3: An RGB-D Visual SLAM Algorithm Fusing Optical Flow Tracking and Feature Matching

In response to the issue of poor real-time performance caused by the large computational load of feature-based methods, this paper proposed an RGB-D visual SLAM algorithm which integrates optical flow and feature-based methods. This algorithm combined optical flow with the feature-based method, effectively improving the real-time performance of the algorithm while maintaining high precision. Firstly, keyframes are matched using feature-based methods, and camera motion is computed by minimizing the reprojection error. For non-keyframes, keypoints are extracted without computing descriptors and tracked using the optical flow pyramid, and pose estimation is achieved by minimizing the photometric error. Secondly, a dual mechanism based on time and space is employed to execute keyframes insertion which ensures tracking stability. Additionally, an adaptive threshold algorithm is proposed to enhance tracking robustness in low-light environments. Experimental results on the TUM dataset indicated that, with comparable average localization accuracy, the proposed algorithm reduces average tracking cost by 37.19%, achieving real-time frame rates of up to 45 fps on mid-to-low-performance CPUs.