In this paper, a two-step approach for vehicles detection is proposed. The first step of approach is to approximate vehicles’ potential locations through searching for shadow area of vehicle low-part. In order to find these shadows, Haar-like feature with Adaboost was used to train a Haar detector offline and the relearning process with hard training samples is applied to increase detection rate. Based on the previous processing, ROI (Region of interest) + HOG + SVM algorithm is used for vehicle verification. At last, K-means approach is used to combine the similar detection results. The experimental results proved that our system could be used for real-time preceding vehicle detection robustly and accurately.