Wavelet-neural network model for automatic traffic incident detection

Abstract

The problem of traffic incident detection can be viewed as a pattern recognition problem. Neural networks are known to solve pattern recognition problems effectively, especially when there is no mathematical model. The computational complexity of neural network algorithms, however, increases exponentially with an increase in the size of the network. Furthermore, with an increase in the size of the network, the size of the training set has to be increased exponentially in order to achieve the same level of accuracy. To overcome this double exponential complexity a hybrid feature extraction algorithm and neural network architecture is created specifically for automatic detection of traffic incidents. The upsteam and downstream traffic data are first filtered by the discrete wavelet transform. Then, a linear discriminant network is used for feature extraction. Finally, the adaptive conjugate gradient learning algorithm of Adeli and Hung is used to train the network.