Video-based driver assistance systems are a key component for intelligent vehicles today. Applications for lane detection, traffic sign recognition, and collision avoidance have been successfully deployed in cars and trucks. State-of-the art algorithms rely on machine learning and therefore depend on invariance conditions, e.g. a fixed image perspective. In order to apply current modules in two-wheeled vehicles one needs to determine the roll angle, i.e. the angle between the road plane and the slanted vehicle. It can either be used for parametrisation of the algorithms or for rotation of the video image back to a horizontal alignment. Using an inertial measurement unit to acquire this data is unreasonably expensive. We propose a video-based module that estimates the current roll angle based on gradient orientation histograms to overcome this flaw. Due to the visual structure of a traffic scene we are able to derive possible roll angles from the gradient statistics by correlation with learnt data. Analogously, we estimate the roll rate by correlating subsequent image statistics and stabilise both measures within a linear Kalman filter. Experiments on real image data from various test scenarios show high accuracy of the proposed approach. Thus, estimating the roll angle / rate from video only, enables us to employ established video-based assistance modules for two-wheeled vehicles without any additional hardware expense.