Abstract: In view of the imaging degradation problem caused by dust, low light, and water mist in coal mine downholes, as well as the shortcomings of the traditional manual duty with long chains and high loads. A coal mine safety hidden danger early warning platform based on "edge cloud collaboration+video intelligent analysis" is designed and developed. The platform innovatively adopts a saliency driven visible light - infrared multi-scale feature fusion method, combined with edge cloud computing power division strategy: the edge end completes video stream processing, image enhancement, and object detection, the cloud end is responsible for rule arrangement, sample reinjection, and model incremental update. The practical application verification of Jinchen Coal Industry indicates that the identification accuracy rate of unsafe behaviors by system personnel reaches 95.6%, with a macro average F₁ value of 0.93; The end-to-end alarm time delay P50 is only 1.2 s, significantly reduced compared to the manual first response (220 s); The closed-loop rate of early warning is improved from 78.0% of manual duty in 2023 to 92.1%; The unplanned shutdown is reduced by about 50 h annually, with a direct economic benefit of 3.44 million yuan. This platform solves the difficult problem of low-quality video identificaition through feature fusion and lightweight inference, optimizes computing power allocation with the help of edge cloud collaboration, and constructs a three-end linkage adaptive early warning closed-loop, effectively improving the coal mine safety management efficiency and economic benefits.