In surveillance, categorical recognition of road vehicles is of great interest. However vehicles are generally textureless. Surveillance videos are typically limited in resolution and quality. The requirement to distinguish similar classes makes the problem even harder. These characteristics make some readily available approaches unsuitable. We argue that in this case shape-relevant models are necessary and rich enough descriptions should be used to discriminate between classes. Based on this, we designed an edge-based feature highly repeatable within one class and discriminative enough to separate different classes. Experiments on our data show that this approach outperforms some state-of-the-art approaches, achieving satisfying high recognition rates.