This paper presents a method of pedestrian detection for automobile applications, based on stereo vision. Vision based pedestrian detection presents a difficulty due to the diversity of appearances. The proposed method overcomes the difficulty mainly through the following two contributions. Firstly, employing four directional features with the classifier increases the robustness against small affine deformation of objects. Secondly, by merging classification and tracking, robustness against temporal change of appearance is improved when considering temporal continuity of classification score. The experiments, performed on approximately 16 minutes of video sequences, confirmed that the method can detect pedestrians with low false detection.