Correction of Selection Bias in Traffic Data by Bayesian Network Data Fusion

Bayesian Data Fusion (BDF) is a well-established method in decision-level fusion to increase the quality of measured data of several equal or different sensors, e.g. [7], [13]. Although the method is powerful, the results of the fusion process are only (1) as good as the sensors are; (2) as good as the a priori knowledge about the sensors is and (3) as good as the a priori knowledge about the underlying process is. For instance, in case of vehicle classification for traffic surveillance by several more or less accurate sensors (item 1), accurate relative frequencies of correct and wrong classifications (phantom detections, incorrect classified vehicles) are required to achieve beneficial fusion results (item 2). This statement is supplemented by an adequate characterisation and quantification of the underlying unknown traffic process (item 3). For an adequate traffic management, there is a particular need for highly accurate traffic data, measured by accurate and reliable sensors, yielding a high degree of acceptance and credibility concerning the significance of the measured traffic parameters. There are a lot of different sensor technologies with different physical functional principles, different performance, problems and thus, differing operational areas [18], [19]. Two currently important coexisting sensor technologies are for instance the inductive loop detectors and video sensors. Loop detectors measure the traffic process temporally, while video sensors enable temporal and wide area measurements, yielding more comprehensive data about the underlying traffic process than loop detectors. Both sensors provide a data quality in accordance with their physical functional principle and in accordance with the influences of the affecting surrounding environment. For instance, an inductive loop detector works properly under fluid traffic conditions, whereas the measurements are not accurate, if there is stop-and-go traffic. Furthermore, vehicle detection and classification may be problematic in case of overtaking procedures, when the loops are overrun only partly, [11], [12]. That means an inductive loop detector is a sensor, which is influenced by the traffic process itself. In contrast to loop detectors, it is a well-known fact, that the most currently employed video sensors usually work poorly under bad weather conditions (e.g. heavy rain, fog, etc.), changing illuminations (e.g. reflections on the road surface) and traffic process dependent problems (e.g. occlusions among the vehicles on the road). Although new methods have recently been developed to overcome the addressed problems [11], the detection errors of currently used video sensors increase to more than 1000%, if the weather and illumination conditions are bad [6]. In contrast, they perform much better (they can reach even the same accuracy as an inductive loop detector), if the conditions for an optimal operation are maintained.