Within the last ten years, there has been a remarkable increased interest in understanding and developing automated traffic surveillance systems, which provide intelligent recognition of target behavior at traffic intersection to efficiently control the traffic signals.

In the real-time automated traffic surveillance system, accurate representation of the traffic scene is crucial for the target recognition in the captured video sequence. However, traffic scene modeling is a difficult task because of the significant intensity variations in the images that are hard to be parameterized analytically. Therefore, statistical learning from the training data has been widely exploited to represent the traffic scene. Moreover, the representation model of the traffic scene is required that it has good generalization performance in real-time traffic application. Recently, support vector regression (SVR) approach becomes a potential statistical learning technique as function estimation of the available training data [1]. SVR estimates not only a function that does well on the training data, but also restrict the class of estimated function by specifying an upper bound on the fraction of training data. It gives the tolerance that new inputs are allowed to lie outside of a distance from the regression estimation.

We describe a background model (traffic scene), a background subtraction process and on-line background update process that we have developed based on the support vector regression (SVR). Following an incremental support vector regression, the proposed background model is on-line updated whenever a training input of intensity values is added to the training set. Such scheme enables the proposed background model to be adaptive to the intensity variance of images caused by the changes of outdoor environment (i.e. illumination changes for visible image and variance of the thermodynamic and atmospheric conditions for infrared image.)