In order to effectively identify the unsafe behavior of underground coal miners, we designed an underground personnel behavior intelligent detection system based on lightweight OpenPose algorithm. The lightweight OpenPose network was used to obtain the coordinates of key points of human skeleton from infrared camera data, and then different recognition algorithms were designed to detect fall, climb and push postures. Experimental results showed that, the algorithm achieved a speed of 30 f/ s and the overall accuracy was 86. 35%. After deploying the detection model to industrial computers and integrating it with alarms, accurate real-time detection and timely alarm notifications for unsafe behaviors was achieved.