Target discrimination in wireless sensor networks remains challenging when sensors have structured electronic noise and deployment settings have variable in-situ clutter. Data-driven learning of discrimination functions is especially hard when deployment sites are remote or hazardous, necessitating reliance on surrogate environments for data collection. The challenge is exacerbated if sensors are resource constrained. We present an intruder discrimination system that addresses these challenges in the context of a battery powered radar mote. The system robustly rejects clutter moving in-situ to trigger discrimination of humans versus other targets only when said targets displace across the scene. Its learning uses a new, generic extension to feature selection methods that leverages cross-environmental robustness instead of a random or bounded model of feature noise. We experimentally validate that our scheme improves the cross-environmental performance medians and dispersions of extant methods by up to 70% and 100% respectively. It achieves optimal performance with very few features given a modest number of diverse training environments, allowing for efficient mote-scale implementation. A mote mesh network has been deployed to detect poachers while rejecting cattle and other non-targets at a rhino reserve in South Africa.

We demonstrate a mote-scale, human-animal classifier based on a micropower radar. Our classifier is automatically learned from diverse data, using features in the joint time-frequency domain. It is being used as part of a wireless sensor network in a forest to create a virtual fence for human and wildlife protection.

This work is on formal modeling, analysis and detection of job interference in large distributed multi-agent systems. Such an analysis usually requires an examination of all the global system states-often impossible due to the well-known state space explosion. We obtain a sufficient condition so that job interference can be detected by observations of individual system component without the knowledge of global system states. Given that the job interference can be detected locally, we propose a guided random walk algorithm for detecting interference. We apply it to Kansei, a large and distributed wireless sensor network system with multi-agents. Ten job interference traces are identified; they have not been detected before by manual analysis and system operations. We further diagnose the detected interference for a correction of system design.

A system for controlling smart sensor networks is described. The system is called the Adaptive Context Information Processing Language (ACIPL) which will allow explicit use of states of context inferred from sensor readings and algorithmic output for distributed control of data fusion in sensor networks. The detailed description of the language including its use for sensor information separation into raw sensor data, feature, objects, and events by the language is described. Furthermore the concept of context which aids the modeling of sensor data with the data fusion hierarchy is introduced, including its types, properties and applications. ACIPL loosely follows the Joint Directors of Laboratories' Data Fusion Hierarchy and provides a powerful tool to facilitate the use of smart sensor networks for the purpose of sensor data fusion.

In following the “Signals of Opportunity” theme of the NAECON '09 Grand Challenge, we explore the use of computer-vision technique proposed in [3]. The proposed approach is not affected by the strength of the microwave, and is more accurate than the conventional time of arrival approaches. Methods based on limited and varying information of markers is discussed. Simple applications and experimentation based on augmenting a motorized wheelchair for vision applications is discussed.