Abstract: Intelligent technology for underground autonomous mobile operation equipment, such as mine trackless rubber-tyred vehicles and rockburst prevention drilling robots, constitutes a crucial component of current intelligent mine construction. Reliable perception of the environment and efficient detection of obstacles by underground autonomous mobile operation equipment are the basis for achieving autonomous movement. Aiming to address the insufficient environmental perception capabilities of autonomous mobile equipment and the low efficiency of multimodal data fusion in complex underground coal mine environments, this paper aims to enhance the accuracy and reliability of obstacle recognition in the operational space of underground mobile equipment. By analyzing the characteristics and limitations of existing identification methods, a coal mine obstacle detection method based on infrared/laser dual-mode point fusion is proposed. Firstly, a simplification method for tunnel point cloud data based on feature point retention is designed. This method preserves rich structural feature information while eliminating outliers and identifying inlier points within the underground spatial plane. Secondly, by incorporating chaotic mapping and opposition-based learning strategies, an Improved Dung Beetle Optimizer algorithm is proposed. This effectively enhances population diversity and search efficiency of the original algorithm, preventing it from becoming trapped in local optima. Subsequently, an infrared image enhancement method based on the Improved Dung Beetle Optimizer is introduced, which effectively mitigates the negative impact of complex scene noise in coal mines on infrared image feature extraction. Finally, an obstacle detection model for coal mine environments is established based on an improved Aggregate View Object Detection (AVOD) architecture. The data input layer of the AVOD architecture was specifically modified to adapt to infrared image features. The effectiveness and applicability of the proposed algorithms were validated through the creation of a dedicated coal mine obstacle perception dataset. Experimental results demonstrate that the proposed method achieves an obstacle identification accuracy exceeding 90% under low-illumination conditions typical of underground coal mine tunnels.