Real-Time SLAM System for Edges Based on Bilateral Filtering and Adaptive Thresholding

This paper presents an edge vision SLAM system based on bilateral filtering and adaptive thresholding. This paper is a complete vision SLAM system including visual odometry, sliding window optimization, and cyclic closure that can be built out of semi-dense maps. At the heart of the approach is the use of bilateral filtering and adaptive thresholding to retain as much edge information as possible and to combine edge images with depth mapping for joint camera pose estimation, with depth mapping further increasing stability in poorly textured environments. The front-end visual odometry module estimates the relative motion between the current frame and the keyframe, using a sliding window to continuously optimise the keyframe, and in the loop-closing and repositioning module, a Fern library of keyframes is built, using a random bag-of-words method to obtain closed-loop candidate frames, closing the loop when a matching candidate frame is found, and adding it to the database otherwise. The method in the paper is experimented on the TUM public dataset, covering a variety of scenes and camera movements, and the experiments show that the system in this paper can effectively construct a semi-dense map of the scene. In addition, the method in this paper performs better in terms of trajectory accuracy for most sequences, indicating that the combination of edge and depth terms selected in the cost function is applicable to a wide range of scenes and that the edge feature-based method has higher accuracy and robustness.