| In the research fields of Intelligent Vehicle, navigation is one of the key approaches. Vision-based navigation system should analyze and process the images obtained from sensors in real-time so as to achieve the purpose of recognition. Facing the high dimensional and large amount of image data, how to extract the effective features has been a vital part for classification and recognition. This thesis focuses on the feature extraction methods for perception images.After analyzing the statistical characteristics of perception images, this thesis makes the images into patches, and conducts research form the color and the texture view to the patch images. From the perspective of color, the thesis analyzes the characteristics of color space, then realizes the RGB space to HSV space conversion, and extracts the color moments features in HSV space. From the perspective of texture, this thesis, on the one hand, realizes FFT-based texture feature extraction methods to extract the image's circular direction and radial characteristics of frequency spectrum; On the other hand, this thesis uses two-dimensional Gabor filtering method to filter the images.After the image feature extraction, the thesis uses the PCA and ICA transformation to reduce the dimension of the feature vector. The new feature vectors keep the second-order and higher-order statistics of perception images.Finally, this thesis studies the real images data collected by the intelligent vehicles, first extracts their color and texture features and uses the SVM classifier method to do the environment classification. Then use the ICA transformation to reduce the dimensionality of the features and compares the classification results before and after. |
