Vehicle identification is one of the frequently studied problems in video surveillance. Commonly, identifying an unknown vehicle object requires a large amount of training instances. Unfortunately, in the large parking scenario, the cost may be prohibitively expensive because of the finitely waiting time from the car owners. In this paper, we show that it is possible to identify a registered vehicle using a single training example. The key insight is that, the human visually representative properties, such as overall appearance and local texture, are described using colors and histogram of oriented gradient (HOG). They are firstly extracted from each patch to form the fundamental representation. After that, the well developed locality-constrained linear coding (LLC) technique is employed to learn a more informative representation. Next, with nearest neighbors (NN) servers as the main classifier, the classical genetic algorithm (GA) is utilized to decide the most informative patches so that the minimum mis-identified error is achieved. The proposed approach is experimentally demonstrated and evaluated on the open Dana36 data set using vehicle images that are taken under an approximate camera pose. Good improvements are obtained with respect to the strategies that do not use LLC encoded features and selectively combined patches.