数字家庭环境中双手交互技术研究

采用基于视觉的双手手势控制方法,构建数字家庭环境。在获取肤色图像后,通过面积过滤和图像方位分辨出左右手的位置,使用形状过滤算法识别出指尖的位置。分析左手静态手势判断所控制的家电,同时跟踪右手指尖移动的路径,通过方向离散化得到方向向量,使用Baum-Welch算法训练方向向量,用Viterbi算法识别动态手势,实现家电状态的控制。实验结果表明,该方法能有效地在数字家庭环境中完成普通家电的控制功能。     

面向语音合成的维吾尔语音素自动切分算法研究

结合维吾尔语语音特征,以建立维吾尔音素语料库为目标,为了减少人工工作量,通过HTK工具实现了音素的自动切分算法:首先完成了文本设计、录音和手动标注等准备工作,设计了上下文属性集,通过训练获得了每个音素的HMM模型,随后对任意输入的语音句子进行了其音素构成部分的自动切分,最后分析了其切分准确度、存在的问题及对策等。实践表明,在语料库的建设中,该研究策略确实节省了大量的时间和人力成本,提高了语音语料库标注信息的一致性和准确性。     

基于AR-连续HMM的故障诊断模型及应用

在状态监测与故障诊断中,被测设备的状态一般不能直接观察到,要通过测量被测设备的表现来感知,这和隐马尔可夫模型(HMM)在本质是相通的。因此可以利用连续高斯密度混合HMM分析被测设备的振动信号,首先以AR模型系数为特征,研究不同状态数与不同混合高斯数对HMM模型分类的影响,再利用较优的状态数与混合高斯数HMM模型进行状态监测和故障诊断,诊断与对比实验结果表明该方法能利用少量样本进行训练和有效诊断。     

Application of the Keyword Recognition in the Network Monitoring

In this paper,the specific application of key words spotting used in the network monitoring is studied,and the keywords spotting is emphasized.The whole monitoring system is divided into two modules:network monitoring and keywords spotting.In the part of network monitoring,this paper adopts a method which is based on ARP spoofing technology to monitor the users’ data,and to obtain the original audio streams.In the part of keywords spotting,the extraction methods of PLP(one of the main characteristic parameters)is studied,and improved feature parameters-PMCC are put forward.Meanwhile,in order to accurately detect syllable,the paper compares the double-threshold method with variance of frequency band method,and use the latter to carry out endpoint detection.Finally,keywords recognition module is built by HMM,and identification results are contrasted under Matlab environment.From the experiment results,a better solution for the application of key words recognition technology in network monitoring is found.     

Multivariate time series anomaly detection: A framework of Hidden Markov Models

In this study, we develop an approach to multivariate time series anomaly detection focused on the transformation of multivariate time series to univariate time series. Several transformation techniques involving Fuzzy C-Means (FCM) clustering and fuzzy integral are studied. In the sequel, a Hidden Markov Model (HMM), one of the commonly encountered statistical methods, is engaged here to detect anomalies in multivariate time series. We construct HMM-based anomaly detectors and in this context compare several transformation methods. A suite of experimental studies along with some comparative analysis is reported.     

基于隐马尔可夫模型的动态手势识别研究

随着虚拟现实技术的飞速发展，人们迫切需要一种自然友好的字符输入方式，于是越来越多的研究人员投入到动态手势的研发当中。本文基于隐马尔可夫模型（HMM）搭建了一套动态手势识别系统。这套系统通过Leap Motion采集动态手势数据，并能够识别36个字母和数字的手势（数字0-9和字母A-Z）。经过大量实验表明，该系统有着很强的鲁棒性，识别单独手势的识别率能够达到93.2%。     

Intent Pattern Recognition of Lower-limb Motion Based on Mechanical Sensors

Based on the regularity nature of lower-limb motion, an intent pattern recognition approach for above-knee prosthesis is proposed in this paper. To remedy the defects of recognizer based on electromyogram (EMG), we develop a pure mechanical sensor architecture for intent pattern recognition of lower-limb motion. The sensor system is composed of an accelerometer, a gyroscope mounted on the prosthetic socket, and two pressure sensors mounted under the sole. To compensate the delay in the control of prosthesis, the signals in the stance phase are used to predict the terrain and speed in the swing phase. Specifically, the intent pattern recognizer utilizes intraclass correlation coefficient (ICC) according to the Cartesian product of walking speed and terrain. Moreover, the sensor data are fused via DempsterShafer's theory. And hidden Markov model (HMM) is used to recognize the realtime motion state with the reference of the prior step. The proposed method can infer the prosthesis user's intent of walking on different terrain, which includes level ground, stair ascent, stair descent, up and down ramp. The experiments demonstrate that the intent pattern recognizer is capable of identifying five typical terrain-modes with the rate of 95.8%. The outcome of this investigation is expected to substantially improve the control performance of powered above-knee prosthesis.      

连续语音识别前端鲁棒性研究

针对大词汇量连续语音识别中识别率不高的问题,提出了将语音增强级联在识别系统前端,在语音增强中将谱减法和对数最小均方误差算法（logmmse）与用于噪声估计的最小控制递归平均算法（imcra）相结合。识别系统使用Mel频率倒谱系数（MFCC）提取特征,用隐马尔科夫模型（HMM）训练与识别。实验结果表明,提出的方法最高能使单词识别率提高38.9%,使句子正确率提高21.8%。该方法用于大词汇量连续语音识别是可行的,有效的。     

基于连续HMM与静态外观信息模型融合的步态识别

利用步态轮廓图像下肢左右边界的距离矢量对步态轮廓图像进行描述,采用步态图像两脚的宽度进行准周期性分析.利用隐马尔可夫模型并融合步态的静态外观信忠模型进行步态时变数据匹配识别.算法在CMU数据库上进行实验取得了较高的正确识别率.