Landmark recognition using projection learning for mobile robot navigation

Mobile robots rely on traffic signs for navigation in outdoor environments. The recognition of these signs using vision is a unique problem. The important aspects of this problem are that the object parameters such as scale and orientation are constantly changing with the motion of the camera. Also, new signs may appear at some time. In this case feature extraction algorithms are unable to meet the constraints of flexibility. Neural networks can be easily programmed for this task. A new learning strategy for self-organizing neural networks is presented. By iteratively subtracting the projection of the winning neuron onto the null space of the input vector, the neuron is progressively made more representative of the input. The convergence properties of the new neural network model are studied. Comparison results with standard Kohonen learning are presented. The performance of the network with respect to training and recognition of traffic signs is studied