Real Time Traffic Monitoring With Bayesian Belief Networks

Modern traffic management systems are, we believe, best implemented as multi-agent systems. When multiple agents have to make decisions on shared knowledge, this knowledge incorporates the uncertainty of underlying information and sensor systems. One approach to deal with uncertainty is the use of probabilistic models called Belief Networks. However, calculating with these models is a NP-hard problem. In order to apply this technology we had to break down its complexity for our specific case. This paper discusses the design choices that we made to boost the performance of our Bayesian belief network and thereby enabling this technique for real-time traffic monitoring in multi-agent systems. INTRODUCTION Traffic management pursues an optimal use of the available infrastructure in terms of traffic flow, safety and quality of life along the road. Over the past decade the number of traffic management instruments has increased rapidly. Ramp metering controls the traffic density and throughput at certain hot spots on the road. Variable message signs can instruct drivers to adapt their driving speed in order to increase safety or lower the level of emissions. The challenge is to select appropriate non-conflicting measures to optimise the quality of the traffic network. Before such decisions can be made, a sound picture of the current traffic situation must be established. Multiple data and information sources are available, but each of them can only partly observe the actual situation and inevitably suffer from some level of inaccuracy. The overall picture inherits this uncertainty, but information systems rarely take this into account explicitly. We envision a hypothesis management system to fuse all sensor data. The method that is discussed in this paper will merge heterogeneous sensor data and other information sources and deal with its uncertainties in a consistent and sound manner. The merging of different types of observations gives the extra opportunity to determine the likelihood of unobserved events that could not have been detected by processing different types of sensors independently from each other. The overall picture that results from this enrichment process will therefore be more reliant and complete. Using this methodology, we want to build a traffic monitoring system to aid decision making for traffic management. The system will be used to recognize the current traffic situation at certain places along a traffic network. The proposed system will function as an intelligent mediator between the sensors and the decision making process.