Forest smoke detection based on deep learning and background modeling
Guohua Wang, Juncong Li, Yongsen Zheng, Long Qi, Wanrong Gu
Abstract
Smoke is an important omen in the early stage of forest fire disaster. However, due to the complexity of outdoor scene, current video-based smoke detection methods are prone to cause false alarms. Aiming on this situation, in order to realize robust real-time forest smoke detection in outdoor scene, a novel method based on deep learning and dynamic background modeling is proposed to suppress false alarm. Firstly, the Single Shot MultiBox Detector (SSD) deep learning network was selected for the preliminary smoke detection. Secondly, taking into account the motion characteristics of the smoke, ViBe dynamic background modeling technology was used to obtain the dynamic region in the video. Thirdly, dynamic region was used to reduce the false alarms of the preliminary smoke detection results. Through massive experiment on various forest real scenes, the accuracy was improved by 30% relative to the method of single SSD, which verified the effectiveness of the method in this paper.