Monocular SLAM Algorithm Based on Improved Depth Map Estimation and Keyframe Selection

Abstract

Monocular simultaneous localization and mapping (SLAM) method can use a simple monocular camera to track the motion pose and generate an environmental map, which is an essential part of modern robot system. Since the monocular camera is hard to directly obtain the image depth information, the depth map is triangulated from the two views. However, the depth estimation is usually not convincing enough through the epipolar line search. Furthermore, the keyframe selection is only based on the Euclidean distance between the two frames, this method is poor real time and robustness. Aiming at the above problems, this paper presents an improved depth map estimation and keyframe selection method for the monocular visual SLAM. For depth mapping, we propose a new photometric error computation method to enhance the gradient of pixel patch in estimated depth maps. For keyframe selection, a method based on the relative motion distance and feature point tracking is used to get a more accurate keyframe. The experiment results indicate that the proposed algorithm can be more accurate in the depth map estimation and keyframe selection.