Citation Context

...ban sensing [11] [10] [1] applications. Although early work [36] mentions the use of microphone as a clue to context, most work found in the literature focus on the camera, accelerometer [6], and GPS =-=[23]-=- as sources of sensor data. There has been significant work on audio analysis and signal processing. The basic problem of sound classification has been as an active area of research [35] [34] [21] [16...

Citation Context

...approach by increasing the detection accuracy of surface-street driving and indoor locations. 3.1 Activity Classification GPS data have been widely used to infer physical activities. Much of the work =-=[16, 17, 18, 19]-=- takes external geoindexes such as map information to add high-level context to location data points and improve the performance of the classifiers. The work of [16, 17] uses GPS data with external kn...

Citation Context

...g different times of day. There is significant amount of work involving location sensors that also aims to extract high-level information about a person’s activities from low-level sensor data. E.g., =-=[12]-=- use information from GPS sensors to construct models of highlevel activity (such as work, leisure, visit) and to identify significant places (such as home, work, store, etc.). Similarly, [11] use loc...

Citation Context

...rge and the scalability of the filter may become a concern. Another possibility is a method based on Conditional Random Fields that learns the significant places visited and transportation modes used =-=[17]-=-. This information is then used to predict future user location. The method relies on detailed parameter learning. Since the sensor is to be selected before actually using it, we need an estimate of s...

Citation Context

...ther differences between our probability and that of P&W. There has been a large amount of very promising work done using Hierarchical Hidden Markov Models (HHMMs)[6], Conditional Random Fields (CRF) =-=[38, 39, 52]-=- and related approaches [27, 42]. These approaches offer many of the efficiency advantages of parsing approaches, but with the additional advantages of incorporating likelihood information and of supp...

Citation Context

...to a general structure in which each child can have multiple parents. Truyen et al. [36] then developed a hierarchical semi-Markov CRF that could be used in partially observable settings. Liao et al. =-=[18]-=- applied hierarchical CRFs to activity recognition but their model requires many GPS traces and is only capable of off-line classification. Wang et al. [37] proposed Dual Hierarchical Dirichlet Proces...

Citation Context

...rs) evolves (the significance of a place may evolve over time). The algorithm to recognize significant locations by combining location trace data with other indicators shares concepts with prior work =-=[49]-=- [47] [66] [65] [63]. The CenceMe approach is distinguished by the way it clusters according to per-user models (rather than globally), and then shares models based on social connections (e.g., presen...

Citation Context

...c visual recognition procedures (Mihailidis et al., 2003). Also, GPS traces of a person can be utilized for activity recognition through a model based on hierarchical conditional random fields (CRF) (=-=Liao, Fox, and Kautz, 2007-=-). However, their model is only capable of off-line classification using several tens of thousands of data points. Our model, in contrast, is capable of on-line classification, which makes it more app...