In order to address the problem of traditional sensors not being able to detect early-stage mine fires promptly, the utilization of the YOLO V5 object detection algorithm for the early detection of mine fires is proposed. The algorithm is optimized by incorporating an attention mechanism, a small target detection layer, and implementing transfer learning. Additionally, precise localization of the fire source is achieved by combining binocular stereo vision. Research findings indicate that the application of the improved YOLO V5 object detection algorithm allows for the rapid and accurate detection of early-stage mine fires. The algorithm's performance is further enhanced by the addition of an attention mechanism, with the most significant improvements observed when the attention mechanism is integrated at the end of the backbone network, in contrast to the backbone network or the C3 module in the neck. Incorporating the ECA attention mechanism into the backbone network proves to be more effective than the CABM attention mechanism, resulting in a 2.2% increase in average accuracy compared to the original network. Experimental testing of the network with the added small target detection layer reveals that, under similar conditions, it can detect a higher number of small targets. Furthermore, the addition of the ECA attention mechanism to the network with the small target detection layer enhances model accuracy by 1.1 percentage points. Building upon the optimization strategy and employing a transfer learning approach leads to an additional increase in model accuracy by 4.8 percentage points. The precise localization of fire sources is achieved by employing an improved YOLO V5 algorithm in combination with binocular stereo vision cameras. The measurement error within a 3-meter range consistently remains below 5%.
 

mine fire detection；YOLO V5；attention mechanism；small target detection；binocular stereo vision