Qualification of traffic data by Bayesian network data fusion

In this paper a method is introduced based on the concept of Bayesian networks (BNs), which is applied to model sensor fusion. Sensors can be characterised as real dynamical systems with specific physical functional principles, allowing to determine the value of a physical state of interest within certain ranges of tolerance. The measurements of the sensors are affected by external, e.g. environmental conditions, and internal conditions, e.g. the physical life of the sensor and its components. These effects can cause selection bias, which yields corrupted data. For this reason, the underlying process, the measurements, the external and internal conditions are considered in the BN model for data fusion. The effectiveness of the approach is underlined on the basis of vehicle classification in traffic surveillance. The results of our simulations show, that the accuracy of the estimates of the vehicle classes is increased by more than 60%.