Error-tolerant Multi-modal Sensor Fusion

Embedded sensor networks (ESNs) are one of the prime candidates for widely used ubiquitous computing systems that will bridge the gap between computing and physical worlds. One of the most important generic ESN tasks is multi-modal sensor fusion, where data from sensors of different modalities are combined in order to obtain better information mapping of the physical world. One of the key prerequisites for all ESN applications, including multimodal sensor fusion, is to ensure that all of the techniques and tools are errorand fault-tolerant while maintaining low cost and low energy consumption. We address the problem of multi-modal sensor fusion (MSF) by developing two generic schemes that are sufficient to solve the MSF problem for a majority of common types of sensors. The first scheme assumes binary sensors; the second considers multilevel sensors. For binary sensors, we have developed a heterogeneous back-up scheme, where one type of resources is substituted with another. For multi-level sensor fusion, we consider a system of sensor readings, where the sensors are of different types. The sensor readings are not completely independent in the sense that the computational part of the system already has a relation model that defines the correlations between different sensor measurements. The multi-level sensor fusion then exploits the correlations between the faulty measurements, and finds the measurement points that minimize the overall error of the model. For each technique, we present efficient algorithms and demonstrate their effectiveness on a set of benchmark examples.