Improved K-means Clustering Color Segmentation for Road Perception

A modified road perception algorithm is presented based on the color image clustering segmentation. According to the comparison of color spaces' uniformity and integrity, an improved K-means clustering algorithm is proposed to segment color images in the space LAB. Firstly, the target area contains road which is gained in images class using the connected domain labeling algorithm. Then, credible road edge points can be obtained in response to alternate-line sampling labeled region of images and assuming the constant of road width consequently. By establishing the B-splines curve model to fit road shape, the algorithm adopts the least square method used to search the optimal control points of splines curve to identify the road boundaries. Introduction Color images [1] [2] are rich in color and texture information, which segmentation is a very complex problem. There are few algorithms that approach or reach the artificial segmentation results. Different color models can be used to represent image color, meanwhile color segmentation algorithms are many different [3]. In order to ensure color images segmentation satisfactorily, not only we must choose an appropriate color space, but also adopt segmentation strategies and methods suitable for this space. Unstructured road and especially field road are provided with complex structure and changing shape[4] [5]. Road models must match the road construction of the actual environment which directly can affect the accuracy of road detection. B-splines curve [6] [7] is flexible and localized Compared to other curves when fitting the actual models. B-splines curve can fit any shape if it is given sufficient control, meanwhile partial modification of the splines will not affect other parts of the curve. Color Models Selection The three components in RGB color space are closely related and hue is sensitivity to the change of component numerical so that RGB color space is not suita le for the K means clustering algorithm. Experiment shows the transformation from YIQ space, YCbCr space into the RGB space is linear, therefore, the correlation of each variable in the YIQ space, YCbCr space is strong, not suitable for the clustering algorithm [8]. Known from other color models conversion relationship, three variables in the LAB color space are retained information of three variables in the RGB color space, but three variables in the HSV color space are retained only part of the variables’ information in the RGB color space. Figure1 is displayed with one road image in the three color space Figure 1(a) is a road image in the RGB model Figure 1(b) is a road image in the HSV model Figure 1(c) is a road image in the LAB model. International Conference on Advances in Mechanical Engineering and Industrial Informatics (AMEII 2015) © 2015. The authors Published by Atlantis Press 1074 (a) RGB model (b) HSV model (c) LAB model Fig.1. The same image is showed in three color space HSV color space can reflect three kinds of features hue、saturation and brightness of the color but it is not uniform color space by perception, namely the change of distance on the space is not entirely representative of the color vision. The LAB model has a wide color gamut, not only contains all the color gamut of RGB, but also solves the problem of unevenly distributed color in the RGB space. So this paper chooses color segmentation of road images in LAB space. Color Image Segmentation We can evaluate the number of categories according to color theory and threshold theory. Firstly we transform the collected road color image into HSV space, and then transform the hue component into a gray image in order to calculate the histogram [9], finally evaluate the number of clustering directly according to the number of the effective wave crest of histogram. The effective wave crest is defined as the number of the area that is more than 1/5 numbers of entire pixels under the curve. If the number of effective wave crest is equal to 2, the number of categories is equal to 3. If the number of effective wave crest is larger than 2, the number of categories is equal to 4. If the number of effective wave crest is less than 2, the number of categories is equal to 2. We take a color image Figure3 (a) as an example and Figure2 is the H component histogram. Because the effective wave number is 3, the number of categories is set to 4. Figure3 (c) ~ (e) are categories images after K-means clustering segmentation. Fig.2. Histogram of hue component The specific steps of improved K-means clustering algorithm are as follows: Step1: Reading a 24 bit true color image (assuming the image size is m×n×3), the color image is converted from RGB color space to LAB color space. Step2: Rank pixels in AB color space into a matrix M (m×n×2) after extraction of AB spatial information in LAB color space. Step3: The data in the matrix is divided into k (k=2, 3, 4, determined by the above method) using k-means clustering algorithm and obtain the cluster center and clustering results automatic assignment of pixels in each category is 1, 2, 3, ... , k). The clustering center is matrix k×2. Step4: Transform the clustering result into a matrix m×n and generate the image as shown in Figure 3(b). Step5: On the basis of step4, the pixels do not belong to the class k the assignment of which is 0, the pixels belong to the class k assignment of which is the original pixel value. Different categories generated from different k images of image are as shown in Figure3 (c) ~ (f). Step6: In different k image, we acquire the optimal solution to obtain the target area.