Pre-processing of inputs to a neural network model for better performance in traffic scene analysis

Abstract

Artificial neural networks, which hold considerable potential for recognising and classifying spatial and temporal patterns, have been used as an efficient method for automatic traffic surveillance, which is an important research topic of intelligent transport systems. An important element in the performance of the neural networks is the composition of the input and output vectors, as well as the network architecture. However, there has been little research on the performance in relation to the attributes of the input and output vectors. In this research, various input vector properties were applied to the backpropagation model, which is the most popular neural network model, to see how the general performance would be affected by the different types of input vector. Experiments were performed with the inclusion of various grey levels, image sizes, edge detection images and combinations of edge and pixel grey information as the input vectors. The experimental results showed that the network performance, in terms of computing cost for training and prediction accuracy, was highly dependent on the characteristics of the input vectors. Two combined input vectors, the grey scale pixels and edge detection image, produced better prediction performance than either the grey values on the pixel or edges alone.