A review of AI vision technology for identifying unsafe behaviors of underground miners

HAO Qinxia; ZHEN Haolong

doi:10.13347/j.cnki.mkaq.20250070

Abstract：The unsafe behavior of underground miners is one of the main factors leading to mine safety accidents,and real-time,ac-curate identification and early warning of such behaviors are critical for improving mine safety management.The rapid advancement of artificial intelligence(AI)vision technology has drawn significant attention to its application in recognizing unsafe behaviors among miners,demonstrating considerable technological advantages.This review aims to systematically explore the progress in AI vision technology for unsafe behavior recognition and to examine its current application status,challenges,and future directions in mine safety management.Unsafe behaviors are categorized into static and dynamic behaviors based on their characteristics to recog-nition methods suitable for different behavior types.The key technical processes involved in behavior recognition,including data ac-quisition,object detection,behavior recognition,and classification,are thoroughly discussed,along with the main methodologies and technological innovations found in current studies.The performance of various algorithms in real-world mining environments is also analyzed.Research shows that AI vision technology has made significant progress in identifying unsafe behaviors of miners and can effectively improve the ability of mine safety supervision.However,it still faces some challenges.Complex factors such as light changes,dust interference,and target occlusion in the mine environment still have a considerable impact on the recognition accuracy,resulting in certain limitations of the existing methods.On the other hand,the lack of data resources has also become an important factor restricting the development of technology.Currently,the available datasets of miners'unsafe behaviors are limited in scale,making it difficult to cover multiple behavior types and complex scenarios,which affects the training effect and practical applicabil-ity of the recognition model.Meanwhile,although some studies have begun to explore multimodal information fusion,this direction is still in the development stage and the related technical system is not yet mature.Future research should further enhance the adapt-ability of AI vision technology in complex mine environments,optimize the algorithm structure to reduce computing costs,and ac-celerate the construction of high-quality and diverse datasets of miners'unsafe behaviors to support more accurate and efficient beha-vior recognition.Meanwhile,the development of multimodal fusion technology is expected to further enhance the robustness of be-havior recognition and provide a more intelligent solution for mine safety management.