APPLICATION OF NEURAL NETWORKS TO VEHICLE RECOGNITION AND TRACKING

Three aspects of the application of neural networks to object recognition and tracking from a moving camera are evaluated on real data. The combination of a correlator with a Kalman filter and a neural network embedded within the tracking loop is shown to increase true recognitions and decrease false recognitions. Large training sets may be reduced by condensing via a k-nearest-neighbour algorithm without significant loss of network performance. Preprocessing the data using a self-adapting layer of discrete wavelet nodes yields performance that is comparable to a multilayer perceptron net.