Visual SLAM keyframe selection method with multiple constraints in underground coal mines

The significant demand of coal mine intelligence has put forward some higher requirements for the intelligent perception of underground mobile robots in coal mines, and the Visual Simultaneous Localization and Mapping (VSLAM) is a key technology for the intelligent perception of coal mine robots. However, due to unstructured environmental features, weak textures, uneven illumination, and small space in underground coal mines, the existing methods that rely on heuristic thresholds for keyframe selection cannot meet the localization and mapping requirements of visual SLAM in underground coal mines. Therefore, a visual SLAM keyframe selection method with multiple constraints in underground coal mines was proposed, which achieves a real-time and robust pose estimation of mobile robot in coal mines and provides data for digital twin in coal mines. Firstly, the proposed method was constrained according to geometric structure, adaptive thresholding was used instead of static heuristic thresholding for keyframe selection to achieve the effectiveness and robustness of keyframe selection. Secondly, the distribution of effective feature points was homogenized by the balance of gravity principle to further ensure the stability of keyframe selection and the denseness and accuracy of created map points. Finally, the steering place was further constrained by using the heading angle threshold to reduce the impact of viewpoint abrupt change on the visual SLAM accuracy. In order to verify the effectiveness of the proposed method, an experimental analysis was conducted in indoor scenes and underground coal mines respectively using an independently designed mobile robot data acquisition platform. Then, the qualitative and quantitative evaluations were made from Absolute Trajectory Error (ATE) and Root Mean Square Error (RMSE). The results show that compared with the heuristic keyframe selection method, the proposed method improves the trajectories RMSE by 29% and 44% in the indoor scenes and the fully enclosed promenade environment, respectively. The proposed method also has a higher robustness, localization accuracy and globally consistent map building effect in underground coal mines.