Inferring Model Structures from Inertial Sensor Data in Distributed Activity Recognition

The Activity-Events-Detectors paradigm describes the rela-tions between activities and sensor nodes under a distributed perspective. The paradigm provides a conceptual abstraction that decouples the full set of activities from the sensor network with the aim of improving the recognition performances and lowering the computational constraints of the detection tasks in the node. In this work, a data-driven methodol-ogy that learns groups of activities and infers the structure of detector models of the nodes of the network under the Activity-Events-Detectors paradigm is proposed. The methodology, defined on a non parametric clustering procedure, makes no assumptions about the number of groups and the relations between detectors and activities: all the relevant in-formation are derived and inferred from the data. Using the inferred structured models, a performance boost of 15% in the final classification accuracy is obtained with a significant reduction of the computational resources of the detectors.