Vehicle detection is a fundamental problem in the research of Intelligent Traffic System (ITS)，specially in urban driving environment. Environment perception for intelligent vehicles depends on the on-board sensors, such as laser, radar and cameras. Laser or radars have shown good performance in measuring relative speed and distance in a highway driving environment. However the accuracy of these systems decreases in the urban traffic environment as more confusion occurs due to factors such as parked vehicles, guardrails, poles and motorcycles. Vision-based approach has the characteristics of wide field of view, relative low cost, good portability and multi-perceptual information. With the improvement of hardware performance (e.g., GPU, DSP) and the computer vision technology, vision-based driving environment perception in real-time can be guaranteed. Vision-based sensor for detecting vehicle can be divided into monocular and binocular system. Mono-vision methods usually utilize different vehicle features to detect targets: appearance-based and motion-based. However, all the monocular methods lost depth information. Without a priori information, we cannot recover the accurate distance from each object in a single image theoretically. Then if the target is occluded, the detection accuracy will decrease significantly. Conversely, stereo-vision method could provide depth information as well as other image appearance information, which is able to improve the detection accuracy obviously. In recent years, there has been significant research dedicated to vehicle detection based on vision. “v-disparity” image is an efficient method to detect the obstacles on the road scene. In many studies, interest points are extracted by one of the stereo cameras, and then using the disparity and depth maps to localize the vehicle. The optical flow is also widely used in the stereo-vision vehicle detection method. In this paper, we propose a hybrid method for vehicle detection based on stereo vision in real-time. In the hypothesis generation (HG) step, we use one of the images from monocular vision to extract vehicle vision features. Firstly, the reliable features will be extracted, such as points, edges and vehicle shadows which are used to determine ROIs. Then, semi-global stereo matching algorithm is utilized to generate disparity maps. Meanwhile, the distance and the real width of the targets can be obtained. Lastly, we exclude the noise ROIs which have larger depth variation in the disparity maps by clustering. In the hypothesis verification (HV) step, we use HOG feature and SVM classifier to verify the final vehicles in the candidate sets. To optimize the system further, we lead the multi-scale classifier and the detection rate increases dramatically. We test our method on two datasets, the KITTI data and ourselves. Experimental results show that the method could achieve real-time performance in vehicle detection. Further, our method can be applied to ADAS on the collision warning system and active braking system for front obstacle detection. Moreover, it can be used as part of the unmanned environment perception system, which has a good prospect of engineering application.