-
ADMZI分布式红外光纤振动传感器原理如图1所示。波长λ1的连续窄线宽光源通过环形器1和PC1后由耦合器1分成两束,然后传播到被测光纤的两个端口。当FUT发生振动时,两束光之间会产生相位差。当光束被同一振动同时调制时,由于位置不同,仍然存在相位差。两束光在耦合2处发生干涉,干涉光通过PC 2、环形器2和DWMD 2后被PD2接收。DWDM2是对光源2引起的后向散射光进行滤波,保证了干涉光的高信噪比。同样波长的λ 2的操作除了光束传播方向相反。其中,λ1、λ2均接近1 550 nm,且λ1≠λ2。该系统具有定位和识别功能,由计算机实现。
图1中,Laser1,Laser2分别为连续波窄线宽激光光束,波长分别为λ1、λ2;DAQ为数据采集卡;Coupler1,Coupler2均是3 dB光纤耦合器;PC1、PC2为偏振控制器;Circulator1,Circulator2为光纤环行器;DWDM1 DWDM2为密集波分复用器;PD2,PD2为光电探测器;L为SOF的长度。
对于事件识别,特征提取是不可避免的,选择合适的特征参数是分类成功的关键。该方法将入侵信号通过EMD分解为IMFS,分别表示不同的平稳信号特征尺度。每个IMF都具有较小的变化和时间特征。峰度特征描述了振动信号的分布,对脉冲信号的微小变化和时间性高度敏感。振动信号的特征向量采用峰度特征。具体提取过程如下。
第一步:计算所有IMFS的峰度特征并进行标准化。
$${T_i} = \frac{1}{n} \times \sum\limits_{k = 1}^n {c_{ik}^4} $$ (1) $$ {T_i}^\prime = \frac{{{T_i}}}{{\sum\limits_{i = 1}^N {{T_i}} }} $$ (2) 式中:Ti为第i个IMF的峰度特征;k为离散点在分量中的位置;n为有限整数;i为IMFS的个数。
第二步:选择包含信号主要特征的j归一化峰度特征,表示为T= [
T'1,T'2,…,T'j]。这些特性将用于对入侵事件进行分类。如图2所示。 -
红外目标识别主要利用红外摄像机采集到的视频图像信息进行入侵事件监控:首先将红外图像进行灰度化处理,经数字采集系统转化成数字量数据,经过上位机对帧图像数据进行优化处理,并采用特征提取算法进行分析,结合行为类数据库对处理结果进行判定,确定是否有如入侵事件发生。其红外目标识别的工作流程如图3所示。
在红外摄像机的可视范围内,一旦有人体目标闯入,读取的图像的灰度差值和背景图像会有较大变化:
$$DI{F_{gr}} = \sum\limits_{i = m}^n {\left| {CU{R_{gr}}(i) - BA{C_{gr}}(i)} \right|} $$ (3) 式中:
$CU{R_{gr}}$ 为监控图像当前灰度信息;$BA{C_{gr}}$ 为背景灰度信息;$DI{F_{gr}}$ 为灰度差值信息。先对差值图像进行小波分解,然后用巴特沃斯低通滤波器对水平和垂直方向高频子带的噪声进行处理,最后进行小波图像的重构。阶巴特沃斯低通滤波器低通模平方函数如下式:
$${\left| {H(u,v)} \right|^2} = \frac{1}{{1 + {{\{ {D_0}/D(u,v)\} }^{2n}}}}$$ (4) 由上式可知,阶巴特沃斯低通滤波器传递函数的一个重要特性是连续衰减的,而理想的低通滤波器传递函数为陡峭,边缘不明显连续。采用阶巴特沃斯低通滤波器处理噪声,降低了图像边缘的模糊度,灰度剧烈变化处不会产生震荡,从而保证了识别信号的质量。
-
支持向量机(SVM)分类是基于决策超平面的思想,在输入空间或高维特征空间中确定决策边界[13-15]。SVM由一组带标记的训练数据集构造线性的函数。这个超平面会把正样本和负样本分开。线性分隔符通常构造为从超平面到最近的负、正样本的最大距离。直观上,这使得训练数据的分类接近但不等于测试数据。不同类型的入侵信号具有不同的特征向量,可以作为RBF的输入。RBF可以根据IMFS的不同特点进行综合分析。
对于多类分类问题,由于输出可以是多个类,并且必须划分为互斥类,因此问题变得更加复杂。MCSVM通常是通过结合多个二进制支持向量机来实现的。文中使用有向无环图(DAG)来获得MCSVM。在训练阶段,构造M×(M-1)/2个二值分类器。在识别阶段,采用有根二进制有向无环图,包括M个节点和M×(M-1)/2个节点。对于一个测试样本,二进制决策函数的求值从根节点开始;然后,它向左或向右移动取决于输出值。
-
实验设置如图5所示。激光光源为1 550 nm分布反馈激光器,强度3.5 mW。外围围栏上的传感电缆总长为2.25 km,采用单模光纤。红外监控采用4 mm无畸变摄像头,像素大小3 μm × 3 μm,采样速率30 帧/s。在基于ADMZI的分布式振动传感器上,通过攀爬、敲击、切割和摆动围栏四类入侵事件进行实验,得到480组数据,图5给出了四种情况下的入侵信号及其特征向量。
图 5 信号及其特征向量。 (a) 攀爬, (b) 攀爬的特征向量, (c) 敲门, (d) 敲门的特征向量,(e) 摆动,(f) 摆动的特征向量, (g) 切割, (h) 切割的特征向量
Figure 5. Signals and their eigenvectors. (a) Climbing, (b) eigenvectors of climbing, (c) knocking, (d) eigenvectors of knocking, (e) waggling, (f) eigenvectors of waggling, (g) cutting, (h) eigenvectors of cutting
将光纤振动信号进行IMFS分解及EMD的预处理,然后利用峰度特征向量通过MCSVM对事件进行分类解析。将SVM结果与红外灰度差值图像像元解析结果进行匹配分析,监控图像输出入侵目标识别结果,如图6所示。
图6(a)中为人为攀爬动作,红外监控比光纤系统能够更为直观准确地识别,并通过矩形框框选入侵目标;图6(b)为较大型动物闯入监控区域时(文中为鸟类),红外监控会发生较高概率的误报,如图b(1)所示;而复合识别方式由于具有多传感器模式识别对比机制,能够很好地杜绝这类误报的发生,如图b(2)所示。
为做进一步验证,笔者进行了120次试验,其中50次试验用于训练。训练结束后,笔者也用这50次训练样本做了一个测试。四种情况的识别率均为100%,证明了该方案的可行性。接下来的测试包括剩余的70次试验,以确定识别效率。复合检测与单一检测方法判别结果如表1所示。
表 1 三种监控方案的实验结果对比
Table 1. Comparisons among experimental results of three monitoring schemes
System scheme Results of the proposed method Results of the misinformation rate Climbing Knocking Waggling Cutting Optical fiber testing 92.3% 90.9% 99.2% 91.6 2.5 Infrared monitoring 59.5% 54.3% 65.7% 62.9% 16.6% Compound detection 95.8% 92.5% 99.8% 93.3% 0.9% 由表1可知,三种对比方法中,复合入侵识别方法相对于单一监测方式有更高的入侵识别率(实验数据为100%)。对于爬越围栏、撞缆、晃动围栏、剪断围栏这四种具体入侵事件,文中方法的识别率分别达到95.8%、92.5%、99.8%、93.3%。更为突出的一点,复合监测具有更低的虚警误报率(实验数据为0.9%),上述结果验证了文中复合方法在具有高效的识别效率的同时,实现更低的误报率,能够更好的地应用在实际监控场合。
Design of composite intrusion detection system based on optical fiber sensor and infrared video
-
摘要: 针对机场、油库等特定区域的高识别率、低误报率入侵事件监控需求,提出了一种基于光纤传感与红外视频的目标识别方法。其中,光纤传感部分采用基于MCSVM的非对称双马赫-曾德尔干涉仪(ADMZI)分布式光纤振动传感器,将EMD(经验模式分解)、将峰度特征与MCSVM相结合以提高识别率;红外识别部分将灰度差值图像通过小波变换提高清晰度。两者经过模式对比算法,实现入侵事件判定。搭建系统做现场实验,结果表明:该方法能够识别四种常见的入侵事件(爬越围栏、敲击电缆、剪断围栏、摇动围栏),平均识别率在92.5%以上,误报率0.9%,相对传统单一传感器方案,该方法在漏报率和虚警率等系统性能上都有较大的改善,能够满足实际应用要求。Abstract: To meet the requirements of intrusion detection with high recognition rate and low false alarm rate in specific areas such as airports and oil depots, a target recognition method based on optical fiber sensing and infrared video was proposed. Among them, the distributed optical fiber vibration sensor based on MCSVM ADMZI (asymmetric dual Mach-Zehnder interferometer) was used in the optical fiber sensing part, which combined the empirical mode decomposition (EMD) and the kurtosis feature with the MCSVM to improve the recognition rate. The infrared recognition part improved the clarity of the gray difference image through the wavelet transform. The intrusion detection was realized by pattern comparison algorithm. The field experiment results show that this method can identify four common intrusion events (climbing fence, tapping cable, cutting fence, shaking fence). The average recognition rate is over 92.5%, and the false alarm rate is 0.9%. Compared with the traditional single sensor scheme, this method has a great improvement in the system performance such as false alarm rate and false alarm rate, and can meet the practical application requirements.
-
5 信号及其特征向量。 (a) 攀爬, (b) 攀爬的特征向量, (c) 敲门, (d) 敲门的特征向量,(e) 摆动,(f) 摆动的特征向量, (g) 切割, (h) 切割的特征向量
5. Signals and their eigenvectors. (a) Climbing, (b) eigenvectors of climbing, (c) knocking, (d) eigenvectors of knocking, (e) waggling, (f) eigenvectors of waggling, (g) cutting, (h) eigenvectors of cutting
表 1 三种监控方案的实验结果对比
Table 1. Comparisons among experimental results of three monitoring schemes
System scheme Results of the proposed method Results of the misinformation rate Climbing Knocking Waggling Cutting Optical fiber testing 92.3% 90.9% 99.2% 91.6 2.5 Infrared monitoring 59.5% 54.3% 65.7% 62.9% 16.6% Compound detection 95.8% 92.5% 99.8% 93.3% 0.9% -
[1] Qin Z, Zhu T, Chen L, et al. High sensitivity distributed vibration sensor based on polarization-maintaining configurations of phase-otdr [J]. IEEE Photon Technol Lett, 2011, 23(15): 1091−1093. doi: 10.1109/LPT.2011.2157337 [2] Ding Z, Steve X, Liu T, et al. Long-range vibration sensor based on correlation analysis of optical frequency-domain reflectometry signals [J]. Opt Exp, 2012, 20(27): 28319−28329. doi: 10.1364/OE.20.028319 [3] 徐宁, 戴明. 分布式光纤温度压力传感器设计[J]. 中国光学, 2015, 8(4): 629−635. Xu Ning, Dai Ming. Design of distributed optical fiber sensor for temperature and pressure measurement [J]. Chinese Optics, 2015, 8(4): 629−635. (in Chinese) [4] 李志辰, 刘琨, 江俊峰, 等. 光纤周界安防系统的高准确度事件识别方法[J]. 红外与激光工程, 2018, 47(9): 0922002. Li Zhichen, Liu Kun, Jiang Junfeng, et al. High accuracy event recognition method for optical fiber perimeter security system [J]. Infrared and Laser Engineering, 2018, 47(9): 0922002. (in Chinese) [5] de Vries J. A low cost fence impact classification system with neural networks[C]// Proceedings of 7th AFRICON Conference in Africa, 2004: 131-136. [6] Yousefi A, Member S, Dibazar A, et al. Intelligent fence intrusion detection system: Detection of intentional fence breaching and recognition of fence climbing[C]//Proceedings of IEEE International Conference on Technologies for Homeland Security, 2008: 620-625. [7] Willett R M, Marcia R F, Nichols J M. Compressed sensing for practical optical imaging systems: atutorial [J]. Optical Engineering, 2011, 50(7): 072601. doi: 10.1117/1.3596602 [8] 王军锋, 盛大军, 刘子畅, 等. 基于微波阵列技术的高铁周界入侵探测系统应用研究[J]. 中国铁路, 2019(2): 41−46. Wang Junfeng, Sheng Dajun, Liu Zichang, et al. Research on application of high-speed railway perimeter intrusion detection system based on microwave array technology [J]. China Railway, 2019(2): 41−46. (in Chinese) [9] 贺宇, 王岭雪, 蔡毅, 等. 折反射周视系统研究进展与展望[J]. 中国光学, 2017, 10(5): 681−698. doi: 10.3788/co.20171005.0681 He Yu, Wang Lingxue, Cai Yi, et al. Research progress and prospects of catadioptric perimetry system [J]. Chinese Optics, 2017, 10(5): 681−698. (in Chinese) doi: 10.3788/co.20171005.0681 [10] 殷笑尘, 付彦辉. 红外/激光共孔径双模导引头光学系统设计[J]. 红外与激光工程, 2015, 44(2): 428−431. doi: 10.3969/j.issn.1007-2276.2015.02.005 Yin Xiaochen, Fu Yanhui. Optical design of common aperture IR/ladar dual -mode imaging seeker [J]. Infrared and Laser Engineering, 2015, 44(2): 428−431. (in Chinese) doi: 10.3969/j.issn.1007-2276.2015.02.005 [11] 仝选悦, 吴冉, 杨新锋, 等. 红外与激光融合目标识别方法[J]. 红外与激光工程, 2018, 47(5): 0526003. Cheng Xuanyue, Wu Ran, Yang Xinfeng, et al. Target recognition method based on infrared and laser fusion [J]. Infrared and Laser Engineering, 2018, 47(5): 0526003. (in Chinese) [12] 范友臣, 赵红莉, 孙华燕, 等. 主被动激光测距探测系统最大距离的计算[J]. 红外与激光工程, 2015, 44(s1): 86−92. Fan Youchen, Zhao Hongli, Sun Huayan, et al. Calculation of maximum range of active and passive laser rangegated detection system [J]. Infrared and Laser Engineering, 2015, 44(s1): 86−92. (in Chinese) [13] 张俊楠, 娄淑琴, 梁生. 基于SVM算法的φ-OTDR分布式光纤扰动传感系统模式识别研究[J]. 红外与激光工程, 2017, 46(4): 0422003. Zhang Junnan, Lou Shuqin, Liang Sheng. Study of pattern recognition based on SVM algorithm for φ - OTDR distributed optical fiber disturbance sensing system [J]. Infrared and Laser Engineering, 2017, 46(4): 0422003. (in Chinese) [14] 李彦, 梁正桃, 李立京, 等. 基于小波和支持向量机的光纤微振动传感器模式识别[J]. 传感器与微系统, 2013, 32(2): 43−45,49. doi: 10.3969/j.issn.1000-9787.2013.02.013 Li Yan, Liang Zhengtao, Li Lijing, et al. Pattern recognition of fiber-optic micro vibration sensor based on wavelet and SVM [J]. Transducer and Microsystem Technologies, 2013, 32(2): 43−45,49. (in Chinese) doi: 10.3969/j.issn.1000-9787.2013.02.013 [15] 张颜, 娄淑琴, 梁生, 等. 基于多特征参量的φ-OTDR分布式光纤扰动传感系统模式识别研究[J]. 中国激光, 2015, 42(11): 1105005. Zhang Yan, Lou Shuqin, Liang Sheng, et al. Study of pattern recognition based on multi-characteristic parameters for φ-OTDR distributed optical fiber sensing system [J]. Chinese Journal of Lasers, 2015, 42(11): 1105005. (in Chinese)