Cross-view object identification using principal color transformation

This paper presents a novel color correction technique for object identification across different cameras. First of all, we project the analyzed object onto the LAB color space and then find its principal color axis through the principal component analysis. Since the L axis corresponds to the intensity, we then rotate the found principal color axis for making it parallel to the L axis. After this rotation, the color distortions among different cameras can be reduced into minimum. Then, a hybrid classifier is designed for classifying objects into different categories even though they are captured under different lighting conditions. Based on a polar coordinate, a sampling technique is then proposed for extracting several important color features from AB plane. Then, using the SVM learning algorithm, a color classifier can be trained for classifying each object into different categories. For the non-color categories, we quantize the RGB channels into different levels. Then, another classifier is obtained for classifying each gray object into its corresponding category. Since the proposed color correction scheme reduce the problem of color distortions into a minimum, each object can be well classified and identified even though they are captured across different cameras and under lighting condition.