ABSTRACT
Accurate estimation of the vehicle ego-motion with regard to the road is a key element for computer vision-based assisted driving systems. In this method, we propose the use of a single camera onboard a road vehicle in order to provide an estimation of its longitudinal velocity by computational means. There are some clear benefits derived from the use of computer vision for ego-motion computation. On the one hand, vision is not subject to slippery, contrary to odometry-based systems. This permits to reduce cumulative errors to some extent. On the other hand, it allows the integration of ego-motion data into other vision-based algorithms for intelligent vehicles, avoiding thus the need for maintaining calibration between different sensors. Some drawbacks must nonetheless be mentioned, such as the small number of feature points normally present in typical road scenes. Conversely, the problem becomes quite the opposite in urban scenes, where really cluttered images must be handled. In this case, the number of feature points increases although most of the information contained in the image is due to outliers. We propose to obtain couples of road features, mainly composed of road markings, as the main source of information for computing vehicle ego-motion. Road markings are normally found in highways and country side roads, where the estimation of vehicle velocity is most useful. Additionally, the use of lane markings allows avoiding the use of complex direct methods [1], [2], [3] for motion estimation. Instead, motion stereo techniques are considered. Motion stereo has great practical advantages as a means for a vehicle to determine its precise distance from external objects. This technique has previously been deployed in the field of indoor robotics [4]. The method is based on sampling a dynamic scene rapidly (e.g., 25 images per second) and measuring the displacements of features relative to each other in the image sequence. Accuracy is another advantage of the method. While the vehicle continues to approach the detected features the accuracy of the measurement improves quickly as the distance decreases. In the sequel an extension of the method for road vehicles and some experimental results will be presented.
