Optimized Vehicle Fire Detection Model Based on Deep Learning

Abstract

Early fire detection is essential to prevent serious problems such as fire disasters and human casualties. These systems should be able to identify fire disasters and send alarms quickly. Existing sensor-based systems tend to be identified after a fire increases because they detect smoke or heat. This paper’s proposed vision-based fire detection system can immediately detect fires from cameras and send alarms faster than sensor-based fire detection systems. To this end, we used deep learning to detect fire in real situations easily. The proposed model can achieve high performance even when catching a vehicle fire by improving the backbone based on YOLOv5. The backbone was enhanced based on Facebook AI Research’s RegNet, and unlike detecting other general objects, it was configured to distinguish and recognize streetlights that can be confused with fires. In addition, various methods such as data enhancement, model ensembling, and transition learning have been added to improve the accuracy of this model. The proposed model has significantly improved by about 11% compared to the average precision of the existing model (mAP).