A Novel Framework for Real-Time Fire Detection in CCTV Videos Using a Hybrid Approach of Motion-Flicker Detection, Colour Detection and YOLOv7

This paper provides a practical view of implementing a real-time vision-based fire detection system that can accurately detect and alert the presence of fires in live CCTV footage. The proposed system combines motion-flicker detection, which uses a background subtraction algorithm to extract the moving objects in a video and a flicker detection algorithm to verify flickering in suspected regions. Then using a YCbCr colour model, the colour of the flickering region is verified to match the characteristics of fire. Further, a YOLOv7 model trained on a large dataset of fire images is employed to verify the results and raises the alarm to alert the presence of fire. Several optimizations have been done to reduce false alarm rate and GPU memory usage times. The accuracy is 98.4%, and false alarm rate is as low as 2.1%.