Research on smoke flame detection based on improved YOLOv3

• Published in: 2023 4th International Conference on Information Science, Parallel and Distributed Systems (ISPDS) pp. 249 - 254
• Main Author: Wang, Lin ke
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.07.2023
• Subjects: Buildings; Deep learning; Fire detection; Fires; Information science; Machine vision; Object detection; Shape; Temperature; Temperature sensors; YOLOv3
• DOI: 10.1109/ISPDS58840.2023.10235414

Recently, forest fires and residential building fires have occurred frequently, which requires automatic fire detection and identification. Although there are detection methods such as temperature and smoke sensors, they are still subject to environmental factors and cannot achieve real-time detection of smoke and flames. In order to solve this problem, this paper proposes an algorithm for pyrotechnic detection and recognition based on improved YOLOv3. First of all, a database containing large-scale pyrotechnic target detection in multiple scenes was constructed, and the categories and positions of smoke and flame were marked separately. Due to the small smoke and flames generated in the early stages of the fire, the problem of insufficient detection of small target objects in YOLOv3 was improved. For detection accuracy and speed, the Yolov3-spp model is improved and used, and a small target detection layer is added. Experiments show that the improved YOLOv3 algorithm studied in this paper can achieve ideal results in smoke and flame detection and recognition in multiple places and complex environmental conditions and can meet the needs of real-time detection.