Vehicle detection plays an important role in the intelligent transportation system. In this paper, we propose a novel approach which performs a spatio-temporal wavelet transform on videos to obtain motion information of vehicles. Different kinds of motion information extracted in sub-bands are combined to form an additive motion image. The additive motion image is then processed by adaptive thresholding, connected components labeling to segment moving region. The moving region is not accurate since the wavelet transform references many consecutive frames along temporal domain. We present a location calibration method to obtain precise moving regions of vehicles. The proposed approach is tolerant to cast shadow and camera vibration. Experimental videos are captured from roadway and expressway. Experimental results show high precision and recall rates.