Abstract: Traditionally, the number of SVM required for classification increases exponentially with the category number. In our approach, some SVMs are omitted by merging several categories into one and dividing a feature space into several subspaces. Combining with the decision-tree classification method, a SVM-based vehicle color recognition approach is proposed by optimizing the number of SVM. In comparison with the nearest neighbor classification method, the proposed approach is superior in recognition speed and accuracy. Experimental results show the proposed solution is an effective multi-category classifier and can be applied to real-time recognition cases.