ATSS-Driven Surface Flame Detection and Extent Evaluation Using Edge Computing on UAVs

The current approach to flame inspection relies mainly on manual methods, resulting in delayed flame detection and inaccurate evaluation of flame extent. In this paper, we propose a surface flame detection model that can be deployed on edge computing devices. The model is based on the ATSS model and has been fine-tuned for this purpose. By equipping a quadcopter drone with the edge computing device carrying this model, real-time flame inspection can be conducted in various environments such as forests and urban areas. Additionally, we introduce a simple and feasible method for flame extent evaluation, which has been overlooked in previous studies. This method utilizes the flame detection results and the camera on the quadcopter drone to calculate the actual area of the flame, providing valuable data support for rescue teams. Experimental results demonstrate that our proposed method surpasses six comparison models, achieving state-of-the-art performance with mAP, AP_50, and AP_75 scores of 0.685, 0.927, and 0.763, respectively. Moreover, the model achieves an FPS of 20.2, meeting the real-time requirements of flame inspection. The proposed flame extent evaluation method accurately estimates the actual area of the flame.