A Study on Forest Flame Recognition of UAV Based on YOLO-V3 Improved Algorithm

In recent years, all regions of China have constantly paid attention to forest fire prevention, which however is still restricted to onsite observation carried out by forest ranger and basic satellite resource survey. The use of UAV system for forest fire monitoring is still in its infancy. To bridge the gap, this study trains the YOLO-V3 algorithm for forest fire detection based on UAV collected data. Traditional flame detection models are commonly based on RGB colors. They can suffer low accuracy and detection speed, and it is still difficult for the YOLO-V3-based model to detect small flames. In this paper, the YOLO-V3 model is improved to support multi-feature detection. Specifically, 208208 smaller resolution feature scales are added to allow the model learning shallow features of flame images. In this way, the learning ability of the proposed model for shallow image information is improved in the feature extraction stage, which can facilitate the dentification of small flame regions. In addition, the prior box is optimized to further improve detection precision. In the experiment, the mAP value can reach 67.6% with detection speed of 190FPS.