The main goal of the project is to provide a low-latency, highly-specific sensor network data delivery to mobile users. Time-critical availability of sensor data is important in providing mobile users with actionable situational awareness. Timely information delivery allows users to respond proactively to situations that arise in their operational space, rather than responding reactively to a central command. The sheer amount of data that is collected in a sensor network, however,makes it impossible for users to pay attention to the majority of the data. Therefore, the need arises to provide users with carefully selected and interpreted data that are relevant to their tasks. Assuming that users operate in the same space as the network, users can act as additional information sources by entering new data, or providing data interpretation to the network. We envision social-network style user collaboration on interpreting sensor data through multi-user data annotations and recommendations.

Research challenges:

• Mobility of users, hard latency constraints and bursty data collection patterns pose significant networking challenges. Existing data delivery techniques in sensor networks, such as, directed diffusion, geographic routing, or collection tree protocols do not optimize for low latency data delivery and experience large delays especially during data bursts. Mobility of data sinks (mobile users) poses additional difficulties for existing protocols -- our preliminary results show that performance of existing protocols greatly degrades with increased mobility, up to the point where the route to a mobile sink cannot be found. We are exploring pre-fetching techniques to allow for timely data delivery as well as lightweight mobility prediction techniques to predict the best gateway to the mobile user in the network.
• Data filtering and collaborative event detection at a sensor nodes, or cluster-head level. Sensors collect large amounts of data, although in many cases, nothing interesting is happening most of the time. Asking users to interpret all this data is not realistic, thus at least basic filtering of data is necessary on the sensor level. We are looking into techniques that would allow to coarsly classify raw data streams (human, vehicle, explosion), to find abstract classiffiers of data streams (flow of people, vehicles, occupancy metric), and to collaboratively detect events (shot, alarm detection, loss of contact, wrong way motion). Cluster-head nodes can provide higher level data classification and additional reliability by fusing data from multiple sensors (building occupancy, shooter location).
• Data matching to users' interests, what we term information brokerage. Even if we can filter-out most of the irrelevant sensor data at the sensor or cluster-head level, we do not expect our resource constrained sensor nodes to provide high level data interpretation. Humans can aid in the interpretation of sensor data, especially for modalities like images, video, or audio that are directly comprehensible by the human senses. However, sending large amounts of data is infeasible in sensor networks, thus most data must stay on the nodes and only highly selected data be sent to users for interpretation. In general, data will not travel long distances -- only those users that are located close to a sensor will interpret data collected at that sensor. Data interpretation is done through data annotations (data tags or labels) which enables interpretations to be transmitted networkwide in a compact way. A lightweight social-network style service running on the sensor network utilizes data annotations for information brokerage and user data recommendation.
• Exploitation of off-line HPC resources. HPC resources at the edge of the network can be utilized to optimize network deployment and performance. In particular, large-scale simulations can be run to predict the best network layouts and optimize network parameter trade-offs. After deployment, data collected during actual network operations can be streamed to HPC resources and used to further tune the network, adpating continuously to changing environmental conditions and traffic load distributions.