Affected by factors such as changes in the coal seam and geological conditions, fluctuations in pump station pressure, and asynchronous actions of the electro-hydraulic control system, the working face often fails to effectively maintain a relatively straight state under production conditions, severely restricting the safe and efficient intelligent unmanned production of the working face. However, the working face straightness detection method based on the high-precision inertial navigation system has high system complexity, high cost, and large maintenance costs, making it difficult to ensure reliable daily engineering applications. In response to the above problems, a systematic study on visual monitoring of the linearity of the working face has been carried out.This method first uses the lightweight improved YOLOv8 semantic segmentation model , which effectively reduces the number of model parameters while realizing the visual structured dynamic division of the working face scene, and accurately extracts the edge features of key target areas such as the base of the hydraulic support and the pushing stroke. Secondly, an optimized dual-parameter image distortion self-correction algorithm is designed. It fully considers the distortion information of key targets on the working face and can obtain the camera's distortion parameters without external calibration. This enables self-correction of distorted images and acquisition of the contour information of key targets on the corrected working face. Finally, a distance mapping model between the visual image pixel plane and the three-dimensional physical space is established to complete the visual estimation of the misalignment distance between adjacent hydraulic supports and the pushing stroke, and based on this, a visual estimation curve of the straightness variation is constructed to achieve intelligent monitoring of the straightness of the working face based on non-contact visual perception. Experimental test verification shows that the lightweight improved YOLOv8 semantic segmentation model achieves real-time dynamic division of visual structure with a 40% reduction in parameters. The average segmentation accuracy of each target area is higher than 95%, and the single-frame image processing time is reduced to 12.3 ms. In addition, the visual estimation errors of the misalignment distance between adjacent hydraulic supports and the pushing stroke are both less than 5 cm, meeting the application requirements of working face straightness intelligent monitoring.