ANALYSIS OF FEEDFORWARD-BACKPROPAGATION NEURAL NETWORKS USED IN-VEHICLE DETECTION

Current vision-based vehicle detection systems use image-processing algorithms to monitor the presence of vehicles on the roads. Recent research has shown that an artificial feedforward neural network tan be trained to provide similar capabilities. A properly trained and configured network should be able to recognize the presence of vehicles in the images it has never been exposed to. This paper discusses the development of a feedforward-backpropagation neural network-based vehicle detection system that recognizes and tracks vehicles with satisfactory reliability and efficiency. Various issues that are important in selecting the optimal neural network model-like the architecture of the network including the number of hidden layers, their units, learning rule, tiling characteristics of the input image and the output representation of the network-are addressed in this paper. This paper also analyzes how the neural network internally learns the mapping knowledge of the input-output training pairs. The final section describes an output post processor that produces the traditional pulse and presence signals.