Robust SLAM through adaptive feature detection and multiscale analysis

Simultaneous localization and map building (SLAM) is the core technology for autonomous localization and navigation of robots using visual sensors in unknown environments. However, monocular vision lacks depth information, which leads to uncertainty of scale information and affects the positioning accuracy. Aiming at the multi-scale problem of monocular vision, this paper proposes a new feature matching optimization algorithm based on orb-slam3 model, and introduces B-spline image pyramid. The algorithm has the characteristics of smooth interpolation, which can make a single image multi-resolution layered, and increase the multi-scale expression while retaining the edge information of the original image; The results are processed by regional local entropy, and the regions with obvious feature information are selected based on uncertainty, which improves the robustness in multi-scale. In this paper, the euroc data set is used for comparative experiments. Although in individual sets, the positioning error is large and the accuracy is not as good as the original algorithm, most experimental results show that the absolute trajectory error (ATE) of the algorithm is better than the traditional orb-slam3 algorithm. The improved orb-slam3 algorithm in this paper provides a high-precision reference model for robot autonomous navigation in complex scenes.