Abstract:  Traditional fire-fighting robots are limited by fire detection and location technology. The detection and location accuracy are greatly affected by the environment, resulting in poor performance, complex deployment and low intellectualization. Against this problem, a set of automatic fire location and tracking system is designed and implemented based on deep learning, which is also integrated with video image processing technology. The system uses a high-real-time deep learning model for fire detection, and it eliminates false alarms by calculating and comparing the structural similarity ratio of the images, as well as combined with the dynamic characteristics of the flame, thereby further improving the detection accuracy. At the same time, the system introduces the secondary detection based on redundant image segmentation to improve the detection rate of small target fire and effectively increase the detection distance of fire-fighting robots. Additionally, our work facilitates the deployment by taking advantage of the monocular camera to locate and track the fire source. Experimental results have demonstrated that the system improves the accuracy and detection distance of flame detection, and has good real-time performance. These results also allow the proposed system to be a prime candidate for fire-fighting robots in some complex environments.

Key words: fire detection, video image, fire source location, automatic tracking, deep learning, fire-fighting robot