In view of the complex traffic and changeable weather and illumination in a scene, traditional vehicle detection method based on the single part model may result in a bad detection. So, using the spatial location relationships existing in multi-components of the vehicle, license plate and rear lamps are selected to construct the probabilistic models, through which vehicles are detected in this paper. In the new method, first, the color image of the road video is decomposed to the rear lamp gray image and license plate gray image through a different color conversion model. After that, the further identification for the license plate is accomplished through the key steps of gradient feature extraction, regional gradient smoothing and local maximum gradient search; similarly, the further identification of rear lamps is accomplished through the key steps of threshold segmentation, connected domain analysis and area calculation. Finally, With the Gaussian Mixture Model, relationships between the parts of the probability are established, and for the relationship model, if it makes the likelihood probability greater than a preset threshold, we argue that these parts belong to the same vehicle, and take the test result as the final vehicle detection result. Experimental results indicate that the new vehicle detection method has a strong adaptability, which can perfectly deal with the bad illumination conditions and target occlusion conditions, as well as a variety of vehicle types.
