An application of neural networks for recognition of traffic marks in the images of wide angle vision sensors with high distortion lens

In our laboratory, we have conducted a research into a special super wide angle lens which is designed to be functionally similar to the human eye. By using this lens we optically obtain foveated information (distorted image). Neural networks are used to make a computer to recognize the real shapes of traffic marks correctly from the distorted image. In this paper, a feature generation method based on discrete cosine transformation is described. The features are used in a backpropagation trained neural networks. We conclude this method can be used in a robot fitted with wide angle vision sensors and the high distortion lens to recognize the traffic makes effectively.