Multisensor based security robot system for intelligent building

Intelligent building can provide safety, convenience, efficiency and entertainment for life in the 21st century. The most importance role of the intelligent building is the security system. We develop a multi sensor-based intelligent security robot (ISR) that is widely employed in intelligent buildings. The intelligent security robot can detect abnormal and dangerous situations and notify users. The robot has the shape of cylinder and its diameter, height, and weight are 50 cm, 130 cm and 100 kg respectively. The function of the ISR contains six parts. There is the software development system; avoiding obstacle and motion planning system, image system, sensor system, remote supervise system and other systems. We develop a multi sensor-based sensor system in the ISR. We use multiple multisensor fusion algorithms to get an exact decision in the detection subsystem of the sensor system. There is an adaptive fusion method, a rule based method, and a statistical signal method. We demonstrate the remote supervisory system to control the ISR using a direct control mode and a behavior control mode. We think that the man–machine interface in a security robot system must have mobility and convenience. Therefore, we use a touch screen to display the system state, and design a general user interface (GUI) to service the user and visitors. The user can remotely control the appliance using a cell phone through a GSM modem, too. The appliance module can feedback reaction results to the user through a cell phone. Finally, we implement the fire detection system in the intelligent security robot (Chung-Cheng-I). If a fire occurs, the intelligent security robot can find out the fire source using the fire detection system. In intruder detection, we program the same scenario to detect the intruder using the intelligent security robot. The intelligent security robot transmits the message of the detection result to the user using a GSM modem for a fire event or intruder, and transmits the detection result to a client computer through the internet.     

Tool support for learning Büchi automata and linear temporal logic

We introduce a graphical interactive tool, named GOAL, that can assist the user in understanding Büchi automata, linear temporal logic, and their relation. Büchi automata and linear temporal logic are closely related and have long served as fundamental building blocks of linear-time model checking. Understanding their relation is instrumental in discovering algorithmic solutions to model checking problems or simply in using those solutions, e.g., specifying a temporal property directly by an automaton rather than a temporal formula so that the property can be verified by an algorithm that operates on automata. One main function of the GOAL tool is translation of a temporal formula into an equivalent Büchi automaton that can be further manipulated visually. The user may edit the resulting automaton, attempting to optimize it, or simply run the automaton on some inputs to get a basic understanding of how it operates. GOAL includes a large number of translation algorithms, most of which support past temporal operators. With the option of viewing the intermediate steps of a translation, the user can quickly grasp how a translation algorithm works. The tool also provides various standard operations and tests on Büchi automata, in particular the equivalence test which is essential for checking if a hand-drawn automaton is correct in the sense that it is equivalent to some intended temporal formula or reference automaton. Several use cases are elaborated to show how these GOAL functions may be combined to facilitate the learning and teaching of Büchi automata and linear temporal logic. This work was partially supported by the National Science Council, Taiwan (R.O.C.) under grants NSC94-2213-E-002-089, NSC95-2221-E-002-127, NSC95-3114-P-001-001-Y02 (iCAST 2006), NSC96-3114-P-001-002-Y (iCAST 2007), and NSC97-2221-E-002-074-MY3.     

面向大规模语料的语言模型研究新进展

N元语言模型是统计机器翻译、信息检索、语音识别等很多自然语言处理研究领域的重要工具.由于扩大训练语料规模和增加元数对于提高系统性能很有帮助,随着可用语料迅速增加,面向大规模训练语料的高元语言模型(如N≥5)的训练和使用成为新的研究热点.介绍了当前这个问题的最新研究进展,包括了集成数据分治、压缩和内存映射的一体化方法,基于随机存取模型的表示方法,以及基于分布式并行体系的语言模型训练与查询方法等几种代表性的方法,展示了它们在统计机器翻译中的性能,并比较了这些方法的优缺点.     

竞争与合作——复杂网络结构化形成的内因

采用博弈理论来探究复杂网络进化过程中结构化形成的内在原因.如今用于研究复杂网络的基本方法有3类:随机图模型(ER模型)、小世界模型、无尺度模型.但这3类方法都没有从根本上解释复杂网络结构化形成的原因,毕竟在现实世界中网络的演化过程并不是一个纯粹的优化过程.考虑到博弈论强调博弈者通过按照既定规则选择不同策略形成各自的博弈结果,认为它是一种用来解释不确定性演化过程的理想工具,而图论又可以很好地表示复杂系统以及系统内部的相互关系和相关属性,故将博弈论和图论结合到一起,对复杂网络进化过程中结构化形成的内在原因进行探讨,得出"参与者之间的竞争和合作是促成复杂网络结构化形成的内因"这个结论.     

一种基于最小生成树的多目标进化算法

怎样保证朝Pareto最优解的方向搜索和如何获得均匀分布且范围广泛的非支配解是多目标进化算法(MOEA)设计时的两个关键问题,它们很大程度上取决于适应度赋值和外部种群维护这两个重要部分.提出了一种基于最小生成树的多目标进化算法(MST_MOEA).在考虑了个体间支配关系的基础上,利用个体与非支配集的距离和不同等级个体的树聚集密度来对适应度赋值;在外部种群的非支配解个数超过规定的种群规模时,用树的度数和树聚集密度对其进行修剪.将其应用于不同维数下9个测试函数,并与NSGA-II,SPEA2进行对比,结果证实了算法良好的收敛性和分布性.     

基于克里金插值的SWAT参数率定大规模并行方法

SWAT（Soil and Water Assessment Tool）水文模型是近几年国内外使用最广泛的分布式水文模型,因其良好的精度、广泛的适应度得到了众多水文工作者的欢迎。但由于其参数率定部分过高的计算量,给水文工作者建模调参过程中造成了大量的不便。为了提高SWAT参数率定的运行效率,本文基于克里金插值理论,对于SWAT原有的优化算法进行改进,提出了一种基于随机过程的参数率定方法,通过一种“期望优化”的方式对拟合曲面进行寻优。并通过Intel MKL和一个对等式并行框架加以并行实现。大大提高了SWAT参数寻优的计算效率以及参数精度,并通过天山北坡流域的模拟验证了此并行模型的有效性,说明了克里金优化方法在水文模拟参数寻优的过程中能够达到良好的精度。     

后验Monte Carlo-Gaussian采样粒子滤波WSN网络定位*

环境因素导致无线传感器网络定位存在噪声影响,实质上是非平滑的非线性问题,针对传统粒子滤波算法在处理该问题时精度不高的缺点,提出一种基于后验泊松分布的Monte Carlo-Gaussian重采样粒子滤波算法的无线传感器网络定位算法。首先,基于粒子滤波算法,借鉴扩展卡尔曼滤波算法采用近似后验高斯分布思想,设计了后验泊松分布Monte Carlo-Gaussian重采样粒子滤波器。其次,采用该滤波器设计实现了无线传感器网络定位算法,解决了非平滑非线性的噪声干扰定位问题。最后,分别对滤波器和定位算法的性能进行了对比仿真实验,结果验证了所提算法的有效性。     

基于低秩表示的非负张量分解算法

为了提高图像分类准确率,提出了一种基于低秩表示的非负张量分解算法。作为压缩感知理论的推广和发展,低秩表示将矩阵的秩作为一种稀疏测度,由于矩阵的秩反映了矩阵的固有特性,所以低秩表示能有效的分析和处理矩阵数据,本文把低秩表示引入到张量模型中,即引入到非负张量分解算法中,进一步扩展非负张量分解算法。实验结果表明,本文所提算法与其他相关算法相比,分类结果较好。     

基于空间纹理相似性的图像角点特征匹配算法

本文针对传统图像角点特征匹配算法的匹配速度慢且准确率低等问题,提出一种基于空间纹理相似性的图像角点特征匹配算法。首先,计算图像目标上角点对应的空间距离矩阵;然后,通过计算图像角点的空间距离矩阵在对应角点邻域LBP特征向量上的瑞利商,将角点在图像灰度特征空间内的度量问题转换为纹理特征空间内幅值的度量问题;最后,根据角点对应的瑞利商的大小,实现不同图像间的角点特征匹配。对不同条件下采集的图像进行角点特征匹配,得到的匹配结果表明本文算法不仅能够很好的适应图像光照、几何变化,得到的匹配正确率较高,同时与传统算法相比本文算法在运行时间上也有大幅度的降低,当处理特征数量较小时平均降低48ms,而匹配特征数量较多时能够降低2408ms。     

The application of a binary division procedure to the classification of forest subcategories using MODIS time-series data during 2000–2010 in China

Forests account for more than 23% of China’s total area. As the most important terrestrial ecosystem, forests have tremendous ecological value. However, it remains difficult to classify forest subcategories at the national scale. In this study, a newly developed binary division procedure was used to categorize forest areas, including their spatiotemporal dynamics, during the period 2000–2010. Time-series images acquired using the Moderate Resolution Imaging Spectroradiometer (MODIS), together with auxiliary data on land use, climate zoning, and topography, were utilized. Hierarchical classification and zoning were combined with remote-sensing auto-classification. Based on the forest extent mask, the state-level forest system was divided into four classes and 18 subcategories. The method achieved an acceptable overall accuracy of 73.1%, based on a comparison to the sample points of China’s fourth forest general survey data set. In 2010, the total forest area was 1.755 × 10⁶ km², and the total area of and shrubs was 4.885 × 10⁵ km². The total area of woodland increased by 2536.25 km² during the decade 2000–2010. The shrub subcategories exhibited almost no change during this time period; however, significant changes in forest area occurred in the mountainous region of Northeast China as well as in the hilly regions of Southern China. The main transformations took place in cold-temperate and temperate mountainous deciduous coniferous forest, subtropical deciduous coniferous forest, subtropical evergreen coniferous forest, and temperate and subtropical deciduous broadleaved mixed forests. The binary division procedure proposed herein can be used not only to rapidly classify more forest subcategories and monitor their dynamic changes, but also to improve the classification accuracy compared with global and national land-cover maps.