Most of the traditional methods, which extract manual feature from data, are based on the particular scene or video source. In this paper, we propose a vehicle detection method that targets to the static images in unrestricted scenes. Firstly, we measure similarities of all initialization regions and merge them by some rules to get bounding boxes. Then the features of these bounding boxes are extracted by deep convolutional neural network (D-CNN) respectively. Finally, Lib-SVM classifier is employed to classify each bounding box and to complete vehicle detection. Compared with traditional method, the proposed strategy performs stronger robustness.