Abstract: Under the background of intelligent mine construction, the application of intelligent equipment has increasingly become the main content of mine intelligent transformation. Intelligent robots in coal mine that are designed for inspecting and dangerous area surveying and doing other tasks depends on the construction of digital map of underground coal mine and the localizing of the robot itself. But most of the traditional localizing methods are inefficient or even ineffective in the underground. Simultaneous Localization and Mapping (SLAM) has become a better choice for underground intelligent robot localization methods. However, due to the high cost of lidar and the poor performance of camera in low illumination environment, it is necessary to design a SLAM localization method that takes into account both low cost and adaptability to low illumination environment. Therefore, a localization method of robot with underground dark light environment adaptability in coal mine is proposed. Firstly, the real images of the gallery of a coal mine in Fengxian County, Baoji City, Shaanxi Province and the dataset of the camera and IMU required for SLAM were collected. According to the images, non-matching dark light and normal light dataset was made, and 3560 images were obtained after data amplification. An EnlightenGAN image enhancement network combined with self-attention module is designed, which takes into account the dependence of different regions of the image without relying on the paired dataset. Based on the ORB–SLAM3 framework, the whole local image detection is introduced to screen the input image, and an improved IMU initialization strategy based on analytical solution is introduced to improve the initialization speed, and the improved image enhancement network is transplanted to enhance the low illumination and uneven illumination images. Experiments on the EuRoC dataset show that the image enhancement-based underground coal mine robot localization method can reduce E_RMS by 12.17% and E_SD by 14.35% in low-light environments. In two actual coal mine roadway scenarios, the low-light environment can be identified, and the increasing number of feature points are extracted and the drift phenomenon of positioning trajectory is reduced by the SLAM system and. Finally, the localizing effect of the system is improved in the dark area of the roadway.