Study Of Applying Manifold Learning To Road Icon Identification

The traffic signs recognition takes an important research area of the intelligence transportation system.The functions of it in the aspects of road safety, information direction and so on will become more and more important. But the current traffic sign recognition in the data amount, recognition rate and other aspects is not so well. Some effective feature extraction methods may lead a good idea to solve this problem. Manifold learning methods can explore the intrinsic relationship between the data, and effectively reduce the dimensions of the data.By giving a research of the typical algorithm of manifold learning in this paper,we experiment on the Swiss Roll data model,use it to campare the linear dimension reduction methods with the nonlinear ones and point out the advantages of locally linear embedding (LLE) methods in the data dimensionality reduction.In order to test and use the LLE dimensionality reduction method to the traffic sign recognition later, first the public face database ORL is used on the face recognition experiment.By comparing the LLE with PCA and LDA methods, LLE method has a good advantage of discovering the low-dimensional features in the image. The advantages and the choice of its parameters through experiments are discussed. Then the directional signs are disposed to verify that LLE method is feasible in the traffic sign recognition applications. In the application research of traffic sign recognition, firstly software framework of the traffic sign recognition is designned. With the color and shape features of road icons, we make use of the moment invariants to detect the traffic signs. Extracting the effective low-dimensional feature of traffic sign images with LLE method and using AdaBoost ensemble learning algorithm with BP neural network classifier for traffic signs classification and recognition are followed. The whole experiments above show that LLE method is feasible and effective for recognition of traffic signs.