An Easy-to-Program Sensor System for Parsing Out Human Activities Dimitrios Lymberopoulos, Andrew Barton-Sweeney, Thiago Teixeira and Andreas Savvides Embedded Networks and Applications Lab, ENALAB Yale Univerisity, New Haven, CT 06520, USA Abstract. We present a new system for interpreting human activity patterns using a sensor network. The paper takes an illustrative approach that presents our methodology using an example, cooking activity detec- tion, in an assisted living application. The activity is detected using a hierarchical probabilistic grammar formulation that is able to detect the activity across multiple instances by reasoning on location information and the kitchen layout map. In our system node level code authoring is automated by a middleware infrastructure that provides a layer of abstraction between the programmer and the sensor nodes. Short de- scriptions of grammars and rules in the form of scripts are converted by the middleware to elaborate node-level code. In our cooking detection implementation, 18 lines of grammar definition are automatically trans- formed into 2600 lines of node level C code that morphs the sensor node into a cooking sensor with binary outputs. 1 Introduction Understanding and acting on the data at the sensor node level would be of immense value to numerous aspects of everyday life. If a small sensor network deployed at home could understand the inhabitant’s behaviors it could provide numerous services. It would be able to guard against unsafe situations, provide warnings, alarms and manage actuation in ways that would greatly improve the quality of life of elders living alone around the globe. To investigate this direction in this paper we present a new methodology for understanding behaviors from low level sensor data using a hierarchy of probabilistic grammars. Our system uses a new imager sensor network to track a person inside a home without tagging the person with a sensor. The collected locations are then correlated with the building map to derive a string of symbols that is processed by a hierarchy of probabilistic grammars to parse out specific behaviors. We illustrate this process with the detection of cooking activity with a single sensor and argue that similar automated classification of activity will enable a new generation of applications in assisted living and aging in place. Our system design is structured to act as an extensible rule-based system classify daily activities using a set of rules and grammar definitions. Using this system