关键词: 复杂地下空间, 图像增强, 多源传感器数据融合, 动态加权, 智能感知

Abstract:

Simultaneous localization and mapping (SLAM) is pivotal for autonomous detection, automatic inspection, and emergency rescue in underground spaces. However, the challenges of narrow and long tunnels, complex terrain, and uneven illumination in underground spaces make LiDAR point cloud and visual image matching highly susceptible to degradation. This, in turn, results in insufficient accuracy or even failure of SLAM when fusing multi-sensor data. To address this challenge, we propose a dynamic weighted fusion SLAM method for multi-source sensor data with enhanced robustness. First, during the visual image preprocessing stage, an image enhancement technique based on the hue, saturation, and value (HSV) color space is employed. This method combines single-parameter homomorphic filtering with contrast limited adaptive histogram equalization (CLAHE) to effectively enhance the brightness and contrast of the image. This improvement strengthens the robustness of visual odometry. Next, the data quality of each sensor is evaluated using a Mahalanobis distance consistency test, which analyzes potential data degradation and adaptively selects the most suitable sensor data for fusion based on the current scene. Finally, considering the key parameters of each sensor, we construct the multi-source sensor factor graph model. The dynamic combination of multi-source sensor data weights is then formed according to the data quality model, allowing for the dynamic adjustment of the weight of each sensor data fusion factor. To verify the effectiveness of the proposed method, experiments were conducted in typical underground spaces such as underground corridors, excavated subway tunnels, and coal mine tunnels using self-designed and integrated mobile robots. Qualitative and quantitative comparative analyses were performed against various state-of-the-art methods. The results demonstrate that the maximum root mean square error (RMSE) of the proposed method is only 0.19 m. The average cloud to cloud (C2C) distance is less than 0.13 m, referencing the point cloud acquired by high-precision terrestrial 3D laser scanning. Additionally, the constructed point cloud maps exhibit superior global consistency and geometric structure authenticity. These findings confirm that the proposed method offers higher accuracy and robustness in complex underground spaces.

Key words: complex underground spaces, image enhancement, multi-source sensor data fusion, dynamic weight, intelligent perception