Abstract: The underground coal mine working environment is characterized by complex spatial structures and significant factors such as widespread equipment vibration. These unique conditions pose severe challenges to the quality of infrared video. Among them, the issues of intra-frame and inter-frame motion blur caused by rapid target movement and the inherent vibration of underground equipment are particularly prominent. Such blurring leads to the loss of image texture, unclear target contours, and degradation of details, thereby reducing the accuracy and effectiveness of subsequent visual processing tasks. To address these challenges, this paper proposes a Transformer-based infrared video deblurring algorithm guided by optical flow with sparse attention. First, optical flow information is utilized to guide the computation of the sparse attention matrix in the Transformer. A hierarchical sparse attention mechanism is applied to adaptively focus on low-contrast key regions, separating infrared features from motion blur components. Simultaneously, recurrent spatiotemporal dependency modeling is introduced to capture and leverage motion information and feature continuity across adjacent frames, addressing the issue of insufficient contextual information in single-frame deblurring. Second, the optical flow estimation is refined through a global motion aggregation module. This module accurately infers and corrects optical flow vectors in occluded areas based on surrounding reliable optical flow information and contextual semantic features, overcoming the limitations of traditional optical flow algorithms in occluded regions, thereby providing more accurate and reliable motion information for the subsequent deblurring process. Finally, a motion blur enhancement method based on physical trajectory modeling is employed to construct a coal mine rock roadway drilling scene dataset with heavy coal dust fog dataset (CMRRD-HCDF, this dataset covers various operational states, interference levels, and complex occlusion scenarios) tailored for heavy dust and fog conditions in coal mine to evaluate the algorithm’s generalization capability in real underground coal mine environments. Experimental results show that the algorithm achieves a peak signal-to-noise ratio (PSNR) of 29.508 2 dB and a structural SIMilarity index (SSIM) of 0.889 3 on the uncooled infrared image deblurring dataset (UIRD), and a PSNR of 33.136 7 dB and an SSIM of 0.958 6 on the CMRRD-HCDF dataset.