Abstract
Traffic computers are widely installed on freeways to harmonize traffic flow and to avoid accidents. This is done by analyzing the current traffic conditions and activating variable speed limit signs whenever reasonable. To fulfill the task the traffic computer needs a lot of informations on the current traffic state being as exact as possible. An important deficiency of these computers is their inability to derive realistic spatial traffic data (e.g., density, trip time) under different traffic conditions and to detect incidents as early as possible. This paper describes the use of new fuzzy logic algorithms to address these problems. The input data for these algorithms do not need any additional data from those available on a traffic computer based on current traffic detection standards. In contrast to known approaches to the use of fuzzy logic for the determination of traffic characteristics, we use data from multiple sensors for faster recognition of changed conditions. The fuzzy rules for the identification of the spatial data have been automatically learned using neural networks. In practice, spatial data are not measurable by today's technology. Therefore, the necessary set of training data for the learning process has been derived by microscopic traffic simulations. Validation of the new algorithms was performed with traffic data from video measurements and by additional traffic simulations. This approach has resulted in much better correspondence between predicted traffic data and observed data on the freeway. This benefit allows a faster detection of traffic incidents in comparison to the algorithms currently used in practice. The main focus of the paper consists in demonstrating the method of automated learning of fuzzy rules using neural nets and in illustrating the importance of traffic simulations for deriving and verifying new algorithms. To exemplify this approach, we concentrate on the determination of spatial traffic density for a specific application in the field. Our results show that we are able to achieve faster recognition of changing traffic conditions than with conventional local computations. The principle behind the described method can easily be applied to other traffic characteristics like trip time or incident probability as well.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have