Abstract: The working environment of mining unmanned vehicles features complex lighting conditions, leading to frequent occurrences of missed detections in pedestrian detection, which undermines the reliability and safety of these vehicles. To address the challenges posed by intricate tunnel lighting conditions, a low-light image enhancement algorithm was proposed. This algorithm decomposed low-light images from the RGB color space into the HSV color space, applied a Logarithm function to enhance the V component, and employed a bilateral filter to reduce noise. Morphological operations were applied to the S component for closing, followed by Gaussian filtering to further eliminate noise. The enhanced image was then transformed back into the RGB color space and subjected to a semi-implicit ROF denoising model for additional noise reduction, resulting in an enhanced image. To tackle issues of missed detections and low accuracy in pedestrian detection, an improved YOLOv3-based pedestrian detection algorithm for mining unmanned vehicles was introduced. This approach replaced the Residual connections in YOLOv3 with densely connected modules to enhance feature map utilization. Additionally, a Slim-neck structure optimized the feature fusion architecture of YOLOv3, facilitating efficient information fusion between feature maps and further improving the detection accuracy for small-target pedestrians, while its unique lightweight convolutional structure enhanced detection speed. Finally, a lightweight convolutional block attention module (CBAM) was integrated to improve attention to object categories and locations, thereby enhancing pedestrian detection accuracy. Experimental results demonstrated that the proposed low-light image enhancement algorithm effectively improved image visibility, making pedestrian textures clearer and achieving better noise suppression. The average precision of the pedestrian detection algorithm for mining unmanned vehicles based on enhanced images reached 95.68%, representing improvements of 2.53%, 6.42%, and 11.77% over YOLOv5, YOLOv3, and a coal mine key position personnel unsafe behavior recognition method based on improved YOLOv7 and ByteTrack, respectively, with a runtime of 29.31 ms. YOLOv3 and a coal mine key position personnel unsafe behavior recognition method based on improved YOLOv7 and ByteTrack experienced missed detections and false positives based on enhanced images, while the proposed pedestrian detection algorithm effectively mitigated these issues.