Abstract: For indoor and outdoor fire prevention, many present algorithms for small target fire detection are lack of accuracy, and can not detect in real time, so an improved YOLOv5 algorithm is proposed. The algorithm widened the number of head layers and introduced selayer layer to accelerate the convergence of classification detection and get more abundant sampling information. The accuracy of the improved algorithm was greatly improved. After the optimization of video stream, the flame could be detected in real time. The experimental results show that the accuracy rate of the improved YOLOv5 model reaches 80.4%, the recall rate reaches 91.3%, and the detection speed reaches 44 frames per second.