Analysis and Design of a Yolo like DNN for Smoke/Fire Detection for Low-cost Embedded Systems

This paper proposes a video-based fire and smoke detection technique to be implemented as antifire surveillance system into low cost and low power single board computer (SBC). Such algorithm is inspired by YOLO (You Only Look once), a real-time state of the art object detector system able to classify and localize several objects into a single camera frame. Our architecture is based in three main segments: Bounding Box Generator, Support Classifier and Alarm Generator. The custom Yolo network was trained using already available dataset from literature and tested with respect to Classical and DL (Deep Learning) algorithms achieving best performance in terms of accuracy, F1, Recall and precision. The proposed technique has been implemented on four low cost embedded platform and compared respect the frame per second that they can achieve in real-time.