Abstract
Keywords: driver assistant system, camera, GPS, sensor fusion, extended Kalman filter
A number of current and future driver assistant systems, such as navigation systems or communication based intersection assistance, are heavily reliant upon precise information as to the vehicle's position on the ground. Detailed analysis of driving tasks has revealed that a number of future driver assistant systems will require accuracy to a resolution of at least 20cm with respect to vehicle position information. Available systems that provide position information, may they be terrestrial, e.g. pseudolites, historical, e.g. odometry sensors, or satellite based, e.g. GPS, are not able to achieve the required level of accuracy. Hence there is a need for new methods which increase the level of accuracy with which the position of the vehicle on the ground is measured. Rather than relying on a single source of data in order to ascertain the vehicle's position, we propose the fusion of data from several sensors, e.g. cameras, inertial sensors, GPS receivers, in order to produce a more accurate position vector. The means of combination is an extended Kalman filter. The extended Kalman filter is described by a vector denoting its current state, for this application this is the estimated position vector, and an error covariance matrix which is a measurement of the estimated accuracy of the state estimate. For the presented application it describes the confidence with which the produced position vector can be trusted. If any degradation in the system's performance should occur, this confidence rating can be used to make the driver aware of the degree of degradation. The costs of the system correlate with the level of accuracy it attains. Hence, an analysis of the confidence can be performed to determine the system parameters which are needed to comply with the accuracy requirements of the task and, thus, keep costs down for applications where a high level of accuracy is less important. The approach is further enhanced by adding a priori knowledge found in digital maps to the sensor data to be combined. Such knowledge comprises the elevation of the terrain, features such as roads, road signs, trees, etc., and any information about constraints (give-ways, traffic lights, one-way systems, etc). Off the shelf sensor components were used to build the system which was installed in a standard production vehicle. The system design and an operational prototype are presented here along with the results of simulation based and real world tests.