An Improved Yolov5s based Real-time Spontaneous Combustion Point Detection Method

The spontaneous combustion of coal pile is very easy to cause fires, especially in the closed coal yards. Early detection and disposal of spontaneous-combustion-point (SCP) could avoid large losses. However, the traditional SCP detection methods are confronted with problem of hysteresis, and cannot effectively detect the small SCP in early stage. This paper proposes a novel detection method based on Yolov5s, by using “FPN+ PAN” in the Neck layer to enhance the SCP expressions. To improve the performance of the model in learning the characteristic of small SCP, the block labeling approach is applied in the labeling process. In addition, TensorRT is adopted to accelerate the forward inference of the model, and the deployment test is carried out on Jetson nano. Through comparative analysis, the proposed method outperforms yolov3s and yolov4s in detecting the SCP in very early stage (smoke or flame), and achieves 7 FPS speed in real-time video stream.