MWIRGas-YOWO: Real-Time Gas Detection via Spatio-Temporal Fusion in Mid-Wave Infrared Video

In industrial production, most potential safety hazards originate from the leakage of combustible gases. Many combustible gases are indistinguishable to the naked eye, and gases have no fixed shape, low contrast, and are colorless. Aiming at the detection difficulties of gases in industrial gas leakage detection, this paper proposes a spatio-temporal fusion detection algorithm MWIRGas-YOWO based on the improved YOWOv2 framework. By establishing a self-built medium-wave infrared dataset in multiple scenarios, it covers gas leakage scenarios under environmental conditions such as different concentrations and different gas pressures in the 3.2-3.4 $\mu \mathrm{m}$ band. Firstly, the input module in the MWIRGas-YOWO network is adjusted to be adapted to the infrared images directly collected by the mediumwave cooled infrared thermal imager. Next, in view of the characteristics of gases having no fixed shape and undergoing diffusive motion, YOWO is transferred to the task of infrared gas detection. Then, by combining the 3D-2D dual-stream feature extraction with the dynamic label assignment strategy, the balance between accuracy and speed is achieved. Finally, the image data of gas leakage in industrial scenarios are actually collected for model training and evaluation comparison. Experiments show that the accuracy rates of the improved model on two videos actually collected at industrial sites reach $78 \%$ and $74 \%$, which is significantly improved compared with mainstream algorithms. This study provides a highly reliable solution for gas detection in industrial production.