Distributed Multimodal Sensing: Sensor Suitability

Abstract

This work presents an algorithm for determination of low-level actions for distributed mobile, heterogeneous multi-sensor systems. The algorithm is based on the definition of a common currency for control systems involving multiple sensors. This common denominator, called sensor suitability, allows the conversion of multiple attributes into a single real-valued number that measures the sensor's performance in relation to a sensing attribute. The concept of sensor suitability facilitates the incorporation of different kinds of sensors and sensing modalities into the system, since decisions are made entirely based on the concept of ?ow suitable a sensor is for the task? independently of the kind of sensor. This results in a flexible and effective way of designing multi-sensor systems for multi-sensor applications such as surveillance, mapping, exploration, recognition, etc. Suitability functions are decision-theoretically sound functions that assign larger values to those conditions that represent the most advantageous sensing condition, thus evaluating the current sensor vantage point. This information is used to position sensors in the same way an expert (i.e., the decision maker) would do it. Multiple attributes are easily combined in an intuitive fashion that allows a human user to aggregate multiple behaviors into a single sensing requirement that otherwise might be too complex or counter-intuitive. Sensor suitability evaluates the sensor's ability to perform the task. However, this is an individual measure that does not convey a sense of how the overall team is doing. Using a special metric, the effect of potential sensor actions is evaluated in terms of their contribution towards the team's objectives, and then those actions that produce the highest collective performance are selected. The algorithm produces adequate performance in sensor teams, while operating under a regime that maintains computational efficiency. Using a one-step approach, the optimal solution is approximated at every step, instead of rigorously computed. The algorithm trades optimality for computational efficiency.