Similarity based Analysis of Networks of Ultra Low Resolution Sensors
Relevance: Pervasive computing, temporal analysis to discover behaviour
Method: MDS, Co-occurrence, HMMs, Agglomerative Clustering, Similarity Analysis
Published: July 2006, Pattern Recognition 39(10) Special Issue on Similarity Based Pattern Recognition
Summary: Unsupervised discovery of structure from activations of very low resolution ambient sensors. Methods for discovering location geometry from movement patterns and behavior in an elevator scheduling scenario
The context of this work is ambient sensing with a large number of simple sensors (1 bit per second giving on-off info). Two tasks ...view middle of the document...
An Adaptive Sensor Mining Framework for Pervasive Computing Applications
Relevance: Pervasive computing, data mining
Method: Data Mining (Frequent/Sequential pattern mining)
Organization: Washington State University
Published: Sensor-KDD 2008
Summary: A mining framework for discovering and adapting to patterns of activity in a smart home (ambient sensing)
The first step is to find frequent and periodic sequences of sensor event patterns. This is done using a modification of the Apriori algorithm to take into account periodicity of patterns. Patterns are post-proccessed so that elements of the pattern that specify context are identified. These patterns are mapped to a heirarchical time structure, that goes from high to low granularity, and each time period has associated with it the patterns discovered in the previous step. The patterns are transformed into a markov chain, and the parameters of the chain are learned over time. A second component, called Pattern Adaptation Module (PAM) allows for monitoring of changes to these patterns. This can either be done by specifying specific patterns to look at (by the user) or automatically. In each case, a measure of the evidence of the pattern being frequent or periodic is maintained and tracked. The performance of the system is demonstrated in a smart home.
Macroscopic Human Behavior Interpretation Using Distributed Imager and Other Sensors
Relevance: Modeling of human behavior, Ambient sensing
Method: Probabilistic Context Free Grammars
Organization: Yale University
Published: October 2008, Proceedings of the IEEE
Summary: Probabilistic context free grammars are used to compose atomic activities into higher level activities. Model is temporally augmented to detect abnormal events
The system is based on a smart home equipped with Imote2 based vision sensors. The goal is to allow designers to specify grammars for behaviors. The 'programming' is therefore by domain experts, who can describe activities with high level scripts. The grammars are analogous to HMM models, however according to the authors, are simpler for users to program. An example grammar can specify for example a normal bathroom visit, which involves, a sequence of toilet use, washing hands and drying them. An incomplete visit would only involve, for example, the toilet use, in which case the system may inform the occupant to complete the other tasks. Time is incorporated by modifying input tokens by user-specified (or learned, in another paper) temporal rules. For example a temporal rule may say that if shower use is greater than a threshold, modify input token to toilet_use_long. The user can also specify for each grammar an execute condition that guides when that grammar would be executed (the problem of segmentation). The paper describes an inference engine for these grammars. System faults can also be detected by specifying grammars for invalid conditions (for example if the user goes from two rooms without...