【作者】 王林泓；

【导师】 龚卫国；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2010， 博士

【摘要】 随着经济的发展与科技的进步,人们对社会公共安全和家居环境安全提出了更高的要求。政府开展的“平安城市”建设,更是将安防工程的建设推向了一个新的高潮。“平安城市”的核心系统包括电视监控系统、电子巡查系统和入侵报警系统等。“平安城市”的建设给安防产业带来了巨大的商机,同时也对各种安防产品提出了更高的技术要求。热释电红外(Pyroelectric Infrared, PIR)探测器作为入侵报警系统中最常见的监控产品之一,它具有功耗低、性能稳定、成本低廉及良好的环境适应性等优点,在家庭、社区和工商业等安防领域有着广泛的应用。但是,现有各种PIR探测器所存在的高误报率的缺点限制了它的应用场合。通过深入的研究发现:虽然PIR探测器自身原理和结构设计存在一定的局限性,但更重要的是缺乏对PIR探测器输出信号的有效分析,没有对不同辐射源的PIR信号进行深入的特征挖掘。因此,将信号处理与模式识别方法引入到PIR信号的分析中,不仅对提升PIR探测器的检测性能具有一定的应用价值,而且对安防系统中一维信号的分析及识别也具有重要的学术意义。本论文在国家“863”计划,国家“十一五”基础研究项目及重庆市科技攻关项目等课题的支持下,针对PIR探测器的高误报率问题,以不同红外辐射源的PIR信号为研究对象,以信号处理和模式识别的理论和方法为手段,深入系统地研究了PIR信号的预处理方法、特征提取方法与特征融合方法,并在此基础上,提出了降低PIR探测器误报率的人体PIR信号的识别算法,为提高PIR探测器的检测性能提供了理论依据和可行的实施方案。本论文主要开展了以下四个方面的探索性研究工作:(1)在深入研究PIR探测器特性的基础上,建立了不同辐射源的等效模型,推导了不同等效模型的有效辐射面积与辐射源位置关系的表达式,分析了人体和非人体PIR信号的差异性。通过仿真得到了PIR探测器的理想输出波形,仿真数据与实际获取的数据具有很好的相似性。这不仅为进一步研究去噪方法提供了可信的“无污染”的原始信号,而且为后续研究和设计PIR探测器提供了有意义的参考信息。最后,验证了PIR信号的非平稳随机性,为研究PIR信号的特征提取方法提供了依据。(2)鉴于人体和非人体PIR信号在时频域上能量分布的差异性,本论文提出了一种基于熵理论的小波包熵PIR信号特征提取方法。小波包分解在频域具有更精细的划分,将Shannon信息熵与小波包分解相结合,可以获取表征PIR信号在时频域中复杂度的特征。研究表明:选择与PIR信号具有相同对称特性的db1小波分解后得到的小波包熵具有较好的分类效果。而人体PIR信号的小波包熵在0-2.5Hz频段显著小于该频段非人体PIR信号的小波包熵值,这表明人体PIR信号的有序性更好。(3)由于实小波变换对PIR信号的数据敏感,即输入数据的变化会对小波系数产生不可预测的结果。因此提出一种基于双密度双树复小波变换(DD-DT CWT)小波熵特征的PIR信号特征提取方法。DD-DT CWT具有良好的平移不变特性,抗混叠性及计算效率高等优点,利用DD-DT CWT小波系数Shannon熵在保留PIR信号近似周期性变化特征的同时,能有效提取人体和非人体PIR信号的时频特征差异,为准确识别不同辐射源提供有效的判别信息。研究表明:4层分解后DD-DT CWT小波系数Shannon熵识别率为87.3%。(4)为了进一步提高识别率,提出了一种基于典型相关分析(CCA)的PIR信号特征融合方法。该方法将两组PIR信号的特征矢量间的相关性特征作为判别信息,既达到了信息融合的目的,又消除了特征之间的信息冗余,为两组PIR信号特征融合后用于分类识别提供了新的途径。研究发现,将PIR信号的全局特征划分为不同的子段,然后再将全局特征与子段特征进行CCA融合,可以获得具有更好分类性能的特征描述。实验研究表明:采用DD-DT CWT小波系数Shannon熵特征与其子段特征进行CCA特征融合后,识别率可达到94.3%,比单独采用该特征识别率提高了7.0%。本论文所提出的3种特征提取及识别方法均能有效地改善现有PIR探测器的检测性能。此外,通过对3种特征提取方法的横向比较发现,基于CCA特征融合的PIR信号识别方法具有最高的人体检测率和最低的误报率。更多还原

【Abstract】 With the development of technology and economy, people put forward higher requirement for the safety of public and home evironment. The Urban Safety Program carried out by government agencies accelerates the development of safe protection engineering. The key parts of the Urban Safety Program consist of television monitoring system, electronic patrol system, human intrusion detection system, and so on. The Urban Safety Program brings great business opportunities, and raises higher technical demands on all kinds of security products at the same time. The pyroelectric infrared (PIR) detectors are most widely used in home and public security system for their low cost, low power consumption, statble performance and excellent environmental adaptability. However, the high false alarm rate of the existing PIR detectors has limited their applications. Based on thoroughly analyses, it’s found out that besides the limitation of their mechanism and structure design, there is no effective analysis and feature extraction for PIR signals of different infrared sources. Therefore, introducing the method of signal processing and pattern recognition into the analysis of PIR signal is not only valuable for improving the performance of PIR detector, but also significant for analyzing one dimension signals in security system.The research proposed in this dissertation is supported by the National High-Tech Research and Development Plan of China, and by the Basic Research Project of the‘Eleventh Five-Year-Plan’of China, and by Key Research Project of the Natural Science Foundation of Chongqing Municipality of China. This dissertation focus on the false alarm rate of the widely used PIR detectors, the study subjects are the PIR signals of different infrared sources collected in different experiments. The preprocessing method, feature extraction method and feature fusion method for PIR signal are studied systematically. The human and non-human recognition mthod for decreasing the false alarm rate has been proposed, so that the practical implementation method for improving the performance is presented.Four main explorative researches on PIR signal recognition are made in this papar:(1) Based on deep research on performances of the PIR detector, the equivalent model of different infrared radiation sources are created, and the relational expressions of effective radiation area and position are derived, and feature differences between human and non-human PIR signal are analyzed. Then the ideal output signal of a PIR detector is simulated. The simulation signal and the collected signal has good similarity, which can supply pure original signal for selecting denoising method, and can provide reference for designing PIR detectors with better performance. At last, the PIR signal is proved to be non-stationary which provides significant reference for feature extraction.(2) In view of the fact there are differences between human and non-human PIR signal in power distribution in time-frequency domain, wavelet packet entroy (WPE) is proposed for feature extraction.Wavelet packet decomposition has finer and adjustable resolution at high frequency bands which extracts more detail features of different infrared sources. Combining the Shannon entropy with the wavelet packet decompaositon, the aquired features descript the complexity of different PIR signals. Experimental results show that db1 wavelet which holds similary symmetry with a PIR signal has best recognition ability. The WPE value of human body is significiantly smaller than that of non-human body in frequency band from 0Hz to 2.5Hz, which demonstrates that the human PIR signal is more orderly.(3) Because the real wavelet decomposition is shift sensitve, that is, small fluctuation will lead to unpredictable results. Double density dual tree complexity wavelet transform (DD-DT CWT) wavelet entropy is proposed for extracting features of PIR signals. DD-DT CWT has good properties of shift-invariant, anti aliasing and high calculation efficiency. Therefore, DD-DT CWT wavelet entropy preserves the properties of approximately periodic of PIR signal, and extracts features which are can be used as discrimination information for different infrared soruces. Experimental results show that the recognition rate is 87.3% when the decomposition level is 4.(4) In order to improve the recognition ability, canonical correlation analysis (CCA) for PIR signal feature fusion is proposed. The proposed method uses correlation features of two groups of feature vectors as effective discriminant information, so it is not only suitable for information fusion, but also eliminates the redundant information among the features. This is a new way to classifiacation and recongiton for PIR signals. Better feature description for classification can be obtained by fusing the local and global features of PIR signals. Experimental results show the recognition rate of DD-DT CWT wavelet entropy fusing with its own sub-pattern based on CCA method can reach to 94.3%, which is 7.0% higher than that of using single feature.The three feature extraction and recognition methods proposed in this dissertation are effective for improving the performance of the existing PIR detector. In addition, by the comparison of the three methods, the CCA feature fusion method has the best human recognition ability and the lowest false alarm rate.更多还原