| Methods for the automatic detection of freeway traffic incidents have been formulated and developed by many authors. Fast and reliable detection of traffic incidents is essential for effective incident management, in which importance is given to the issues of reducing traffic delay, increasing traffic safety, and minimizing traffic congestion.; While existing automatic detection techniques provide certain necessary information for incident management, they suffer from a high level of false alarms and prolonged detection time delay. A more reliable technique is needed to attain an improved incident management.; It is shown in the present study that neural networks with the cascade correlation architecture can be used to achieve better performance in detecting freeway incidents. The application of the cascade correlation algorithm to incident detection problem promises several advantages. First, the algorithm not only eliminates the need to guess the appropriate size of the neural network, but also optimizes the network, in terms of the number of hidden units, by itself. Secondly, at any given time, we train only one layer of weights in the network. The rest of the network is not changing.; In order to improve accuracy of the cascade correlation algorithm for incident detection, three multi-network models are introduced. It is shown that, using a hierarchy of primary and secondary neural networks, a multi-network architecture can perform better, in terms of false alarm rates, than a single neural network. An incrementally trained neural network algorithm is also devised in order to improve the accuracy of incident detection. |
