Research On Feature Extraction Of Vehicle And Classifier Design

Pattern dection system contains two important parts. They are feature extraction and classifier design. Using effective feature extraction method to process data and applying the artificial neural networks to classification of vehicles dection are one of the effective spproaches. two feature extraction theory which have been used to get feature of vehicle waveform and the application of two classification methods which are artificial neural networks and Minimum distance classification are further discussed in this dissertation.The simulation results show that the proposed feature extraction method and classifier have good effect.Component analysis is a statistical analysis technique which can reduce the dimension of variables. Component analysis reduce the dimension of data, and elim inate the data correlation with retaining the most information. The KPCA(Kernel Component Analysis) is in fact an improved component analysis algorithm, which transform the original space into a new high-dimension space. In this new high- dimensional space, it can obtain main component by the use of kernel tricks.Traditional pattern dection system usually makes use of relativity classifier or distance between the sample to classify vectors, but this system can't achieve good classification results when the input samples contain noise or the number of categories too is large. Adaptive Resonance Theory network have some good characteristics which other classification methods do not, and this network can resolve the problem of plasticity and stability, can quickly and steadily identify the objects which have been studied and can quickly learn the objects which have been not studied. So we can use its stable merits and competitive learning to classify features. Traditional ART network has been improved in two aspects, one is the cosine of the angle between the introduction of the minimum standards as an input pattern, the other is the introduction of identification criteria for judging similarity. So Simplified ART network learning algorithm has been established and it is suitable for the classification of high- dimensional feature value target. Then a new classifier is established based on that network. Experimental results on the data obtained by Component Analysis and Kernel Component Analysis show its high dection effect, dection rate and self- adaptation which far exceed other neural networks.