【作者】 王衡； 蒋斌； 岳玮宁；

【Author】 WANG Heng, JIANG Bin, YUE Wei-Ning+ (Department of Computer Science & Technology, Peking University, Beijing 100871, China)

【机构】 北京大学计算机科学技术系；

【摘要】 上下文感知的人机交互对于提高计算系统,特别是普适计算系统,的智能性具有关键的作用.本文在我们提出的分层上下文感知交互结构的基础上对上下文传感器底层数据的融合进行了着重讨论.为了满足实时性要求并避免噪音数据干扰,通常使用的数据融合算法是Kohonen自组织图(KSOM).本文对KSOM算法进行了深入分析,并提出了通过K均值算法改进KSOM的数据融合方法.在智能导游系统TGH中的应用表明,本方法比原始具有更好的效率和融合效果.更多还原

【Abstract】 Observing communication between humans we can see that the action of a person is always performed in a certain situation and lots of information is implicitly exploited in the exchange of messages. What happens in the surrounding environment often supplies valuable information that is vital for the communications. If computer applications can also utilize such information to characterize the situation, the activity and adaptability of interaction will be enhanced. In other words, computers should be able to have a certain understanding of user’s behavior and surrounding states in a given situation, and use this knowledge, named context, as additional input. The hints carried in context could help ubicomp applications select the most appropriate mode and automate tasks, so that the attention minimization can be obtained. Context is an abstract concept and therefore difficult to capture directly. Salber and Dey’s context toolkit and Schmidt’s architecture proposed the idea of layered abstraction to extract ambiguous environmental situations into executable contextual information. The majority of successive works derive from this idea, including our hierarchical architecture: Physical or logical sensors are used to capture raw data from environment. Then the raw data is divided into several basic elements, which provide an abstraction of sensor data. Generally, each element is dependent on one single sensor; but using the data of one sensor, multiple cues can be calculated. Finally, a clustering algorithm is used to cluster cues into contexts. In this paper we focus on how to cluster sensor data into contexts. Mapping the sensor data to contexts is a difficult problem involving near real-time classification and training of patterns out of noisy sensor signals. In general, Kohonen’s Self-Organizing Maps (KSOM) is used for this purpose. This article discusses the KSOM algorithm, and analyzes its advantage and disadvantage. Finally, it proposes an adaptive approach that uses a Self-Organizing Map enhanced by K-means clustering algorithm for clustering sensor data into contexts.更多还原