Smart fire alarm system with person detection and thermal camera

Abstract

Fire alarm is crucial for safety of life and property in many scenes. A good fire alarm system should be small-sized, low-cost and effective to prevent fire accidents from happening. In this paper we introduce a smart fire alarm system used in kitchen as a representative scenario. The system captures both thermal and optical videos for temperature monitoring and person detection, which are further used to predict potential fire accident and avoid false alarm. Thermal videos are used to record the temperature change in region-of-interests, for example, cookware. YOLOv3-tiny algorithm is modified for person detection and can be iteratively improved with the hard examples gathered by the system. To implement the system on an edge device instead of a server, we propose a high-efficiency neural network inference computing framework called TuringNN. Comprehensive rules enable the system to appropriately respond to different situations. The proposed system has been proved effective in both experiments and numerous cases in complex practical applications.