Semi-direct Visual SLAM Algorithm Based on Improved Keyframe Selection

In order to meet the continuous improvement of real-time performance, positioning accuracy and robustness for visual Simultaneous Localization and Mapping (SLAM), and to address the problem that sparse maps are difficult to directly use for autonomous navigation, a semi-direct visual SLAM algorithm based on improved keyframe selection is proposed. It is based on ORB-SLAM2 framework. The front-end fuses feature-based method and direct method to get the camera initial poses. By minimizing the re-projection error to optimize the pose, it obtains a more accurate pose estimation. The proposed redundant keyframe deletion algorithm is applied in the back-end. It is based on the relative motion distance between keyframes, so as to delete the keyframes with overlapping visual fields and reduce the number of keyframes. The filtered keyframes are used for closed-loop detection and dense point cloud map construction. Then the map is filtered and converted into an octree map with higher storage efficiency. Experiments show that the algorithm can improve the real-time performance while ensuring the positioning accuracy, and has strong robustness to the sparse texture or camera’s fast-moving scene, and it can reconstruct the 3D maps of different environment.