Dickmanns and Zapp 1987 : A visual feedback control system has been developed which is able to guide road vehicles on well structured roads at high speeds. The road boundary markings are tracked by a multiprocessor image processing system using contour correlation and curvature models together with the laws of perspective projection. Feature position data are the input into Kalman filters to estimate both the vehicle state vector relative to the driving lane and road curvature parameters. Velocity is measured conventionally. Longitudinal control by throttle and braking is geared to lateral acceleration due to road curvature; lateral control has an anticipatory feed-forward and a compensatory feedback component. The control system has been tested with a CCD TV-camera and image sequence processing hardware in a real-time simulation loop and with our experimental vehicle, a 5 ton van equipped with sensors, onboard computers and actuators for autonomous driving.

1. INTRODUCTION
With the substitution of animals towing carriages by engines on the vehicle, engineering opened up a new aera of transportation. On the one hand, this brought about new ranges of speed and mobility but on the other hand it necessitated the continuous attention of a driver to control the vehicle. A horse is able to learn the way it is being guided (e.g. in the middle of the right half of a road) and the road system around its home; so it is able to at least partially navigate correctly in a known environment. In addition, trained horses can take oral commands like which way to take at a road junction. The (of course limited) autonomous capabilities, which the trained animal provided using its visual and auditory senses and the brain behind it have been traded for the performance advantages of the auto-mobile.
   The highly integrated electronic devices of the near future may allow regaining and even improving the sacrificed capabilities for modern high performance vehicles. High frequency image sequences (like TV-signals) and auditory inputs can be analyzed by computers to allow autonomous vehicle guidance. At present, this is possible only in a simple way but this will change with the highly parallel VLSI systems to come.
Investigations into autonomous vehicle guidance have been performed for some time. Systems using special installations along the road like beacons or buried wires are not considered here; they do not allow to detect obstacles, neither near the vehicle nor in anticipation along the road. This capability to detect the actual state of the environment and to check the viability of locomotion is considered to be a main ingredient for autonomous driving in the sense adopted here.
   The planetary rover of NASA (Gennery, 1977) probably was the first project to spur serious research in this field for some time. Most of the other systems investigated in the area were laboratory carts of small size and low speed that were able to move only intermittently, either measuring or moving (for a survey see [Giralt 1984]). Pioneering work in this field has been done by Moravec [Moravec, 1980]. A new drive towards real applications originated from the US-DARPA Program on Strategic Computing, in which the Autonomous Land Vehicle (ALV) has been selected as one of the demonstrator projects. Road following at low speeds was defined as an initial goal and has been achieved in 1985 [AWS, 1986]; the goal of the project is to achieve the capability of autonomous cross-country mobility.
   In contrast, the research reported on here aims at providing the capability of autonomously guiding a high speed vehicle along a well structured, normal, limited-access highway with one way traffic, no crossings and a limited class of participants (e.g. Autobahn). This puts relatively low requirements on image processing and yet is possibly of practical importance (autopilot). With the advent of more powerful miniaturized computer systems this may even become economically viable for general traffic applications. Future growth potential exists in the direction of autonomous mobility in more complex environments like state-, country- and eventually city road nets. The system is intended to fit into the present traffic situation with its rules and regulations, requiring no additional installations along the road network and allowing a smooth, gradual deployment (like autopilots in aviation). Like these it is conceived as an add on option to driver control systems.