Background Precise positioning of tunnel boring machines (TBMs) in underground coal mines plays a fundamental role in the automated and intelligent guidance and control of fully mechanized heading face. However, traditional visual positioning methods show limited application effects in underground roadways due to their narrow and enclosed spaces, insufficient illumination, and sparse textures. This study proposed a visual positioning method for TBMs in underground coal mines based on anchor net features.

Methods A three-stream depthwise separable convolutional neural network (TSCR-NET) for image enhancement was employed to estimate the reflection, illumination, and noise in images individually. Through illumination adjustment while suppressing noise, images with uniform illumination and clear textures were obtained. This contributed to enhanced adaptability of the visual positioning system under complex illumination conditions. An extraction and matching method for anchor net line features was designed. This method enhanced the extraction capacity using the edge drawing lines (EDLines) with adaptive thresholding and improved the matching accuracy using the structural similarity index measure (SSIM). A pose estimation model with minimized reprojection errors of line features was constructed. In combination with pose graph optimization, this model enabled precise TBM positioning. Furthermore, an experimental platform was established. Accordingly, experiments were designed for quantitative analyses of image enhancement, line feature processing, and positioning performance.

Results and Conclusions The results indicate that the TSCR-NET yielded higher peak signal-to-noise ratio (PSNR) and SSIM values compared to the multi-scale retinex with color restoration (MSRCR) and zero-reference deep curve estimation (Zero-DCE) algorithms. The line feature processing method designed in this study outperformed traditional algorithms in the quantity of extracted features and matching accuracy, laying a solid foundation for subsequent positioning processes. In terms of positioning experiments, the method proposed in this study was compared to other line feature-based visual positioning methods under the EuRoC dataset and the real roadway scene. The comparison results revealed that the proposed method outperformed the real-time monocular visual SLAM with points and lines (PL-VINS) algorithm under nine EuRoC data sequences. Furthermore, in an anchor net-supported roadway scene, continuous TBM tracking was conducted within a range of 60 m. The proposed method yielded a maximum error of 163 mm, indicating a 23.5% reduction compared to the 213 mm obtained using the PL-VINS algorithm. Additionally, the root mean square error (RMSE) decreased from 0.531 to 0.426, suggesting a reduction of 19.8%. Overall, the visual positioning method proposed in this study enjoys high accuracy and stability, providing a valuable reference for long-distance pose detection of TBMs in underground anchor net-supported roadways.