A Fire Monitoring and Alarm System Based on YOLOv3 with OHEM

Abstract

Fire monitoring and alarm has been paid close attention since fire could cause irreversible losses of people's lives and properties, especially in important places, such as substations and hospitals. Currently, most fire monitoring and alarm systems depend on sensors which have such disadvantages including low sensitivity, slow response, small coverage and poor stability. In this paper, a fire monitoring and alarm system based on video surveillance system, which is widely equipped in public places, is presented. Compared to sensor-based fire detection, the proposed system has advantages of quick response, insensitiveness to environmental temperature and accompanying images of surveillance scenes. The proposed system depends on a fire detector and a smoke detector based on YOLOv3. Due to the lack of public data set, a new data set was established by collecting fire and smoke images from internet and labelled by Labellmg. Fire detector and smoke detector are trained on fire data set and smoke data set respectively. Furthermore, online hard example mining(OHEM) is adopted to deal with the imbalance between simple samples and hard samples in the constructed data set. In the detection process, results of fire detector and smoke detector are combined to distinguish whether there is a fire alarm or not. Experimental results on homemade data set demonstrated the effectiveness of the proposed algorithm.