Camera-Based Navigation: Convolutional Neural Networks vs Rule-Based Approaches

We compare the performance of rule-based image processing algorithms for autonomous navigation with that of a Convolutional Neural Network approach based on an Nvidia CNN architecture. We detail our implementation of Hough-Transform-based and pixel-based navigation algorithms and discuss their advantages and disadvantages, including a quantitative comparison of their navigation performance in terms of speed, stability and reliability. We describe how rule-based algorithms can be used to train Convolutional Neural Networks that result in enhanced navigation with high speed, stability and reliability. We discuss the advantages of training Convolutional Neural Networks with binary images instead of colour images and the resulting enhanced performance. We present our experiments on autonomous navigation with model cars racing on unknown tracks using the different algorithms and conclude with some recommendations for the use of each algorithm.