Deep Learning-Powered Vision System for Fire Detection and Localization in Harsh Environments

Fire is one of the most dangerous mishaps that can occur in homes, schools, workplaces, businesses, and forests. It can result in significant losses, casualties, and major equipment damage. In order to protect ecosystems from fire disasters, advanced disaster response procedures must be put in place. Traditional sensor-based fire detection systems cannot provide an alarm until the heat or smoke reaches the sensors. Therefore, it is evident that a rapid, robust, and reliable system capable of detecting fire at an early stage must be developed. Surveillance cameras used for security purposes can be employed to detect fires in structures. This paper aims to establish two fire detection techniques implemented using Colab. The first technique depends on computer vision and uses color models such as Hue Saturation and value (HSV) and YCbCr. The HSV and YCbCr models perform well in terms of fire detection. The second method relies on deep learning and utilizes object detection models such as YOLOv8 and YOLOv5. Two different versions of YOLOv8 and YOLOv5 are used in this research, namely YOLOn and YOLOm. YOLOv8n completed the training in 196 minutes and the average time for a single inference was 70 ms while YOLOv8m took a training time of 289 minutes and required an average of 355 ms for a single inference. However, as can be seen from the table, a lower recall, precision, mAP50, mAP50-95 was achieved but, interestingly, it also had lower losses and Interestingly, You Look Only Once V5 n was trained in 179 mins with average inference time of 55 milliseconds while YOLOv5m was trained in 249 mins with average inference time of 254 milliseconds but passed through a smaller number of losses than YOLOv5n.