摘要
针对复杂场景的烟雾检测准确性低等问题,提出了一种基于多种纹理特征的烟雾检测算法。首先,为了提取出完整的烟雾前景区域,在背景建模时融合了视频像素点的时间和空间信息。然后,在研究和改进局部二值图特征的基础上,提出了3种新的具有高辨别力和鲁棒性的纹理特征,分别为梯度局部二值图特征、多量级局部二值图特征以及局部共生二值图特征。通过提取前景区域局部图像块的这3种纹理特征,利用支持向量机分类器进行分类。最终,通过对3种纹理特征的综合决策检测出准确的烟雾区域。在烟雾图像数据库的测试下,该算法的平均检测出率、误报率及错误率分别为0.978、0.014及0.016,与现有最优算法相比,性能分别提高了0.6%、0.97%、0.83%。大量视频实验结果表明,该算法对复杂场景适应性强,检测准确率高,对比现有视频烟雾检测算法检测率提高了2%~4%。
A novel smoke detection algorithm based on the multi-texture features is proposed to solve the problem of low detection rate of smoke in complex scenes. In order to extract the complete smoke foreground area, both temporal and spatial information of the pixels are fused in the background modelling process. Three novel discriminative and robust texture features are proposed by carefully studying and improving the local binary pattern feature, and are further utilized for support vector machine training in the foreground patch area. Finally, the accurate smoke area is detected through a comprehensive decision making on these features. Test results on smoke image data sets show that the average detection rate, false alarm rate and error rate of the proposed algorithm are 0.978, 0.145 and 0.162, respectively, which gain improvements of 0.6%、0.97% and 0.83%, respectively, compared with the existing optimal algorithm. Extensive experiments on challenging scenes show that the proposed algorithm outperforms other video-based smoke detection methods by 2%- 4% in the detection rate.
作者
赵敏
张为
王鑫
刘艳艳
HAO Min 1, ZHANG Wei 1, WANG Xin 2, LIU Yanyan 3(1. School of Microelectronics, Tianjin University, Tianjin 300072, China; 2. Tianjin Fire Research Institute of MPS, Tianjin 300381, China; 3. College of Electronic Information and Optical Engineering, Nankai University, Tianjin 300071, Chin)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2018年第8期67-73,共7页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(61474080)
公安部技术研究计划竞争性遴选项目(2016JSYJD04-03)
关键词
烟雾检测
时空背景建模
纹理特征
支持向量机
smoke detection
spatio-temporal model
texture feature
support vector machine