Stationary or low-speed vehicles and fallen objects from vehicles on roads may cause serious accidents that involve following vehicles. If AHS (Advanced Cruise-Assist Highway Systems) are introduced, which recognizes such obstacles using sensor installed on a road and warns following drivers, the accidents would be prevented. Experiments have been performed to detect obstacles on roads using an image sensor and radar installed on the roadside. The image sensor, which utilizes a light wave, is widely used in traffic monitoring because it is a very effective detector. However, the misdetection rate of this sensor increases with changes in illumination. On the other hand, the detection performance of radar is not affected by changes in illuminance. Radar is affected by multipath interference, but this does not affect the detection performance of the image sensor. Therefore, by using radar together with an image sensor, it is possible to achieve stable vehicle detection. A stationary or low-speed vehicle can be recognized by tracking technique based on the Kalman filter. Vehicle tracking is effective for detection of obstacles except vehicles. In the conventional vehicle tracking method, the two dimensional state vector of the vehicle is estimated. In this method, range and azimuth measured by radar, and elevation and azimuth measured by an image sensor are transformed into a two dimensional position on the road plane. This is done under the assumption that the road surface is flat. Therefore, tracking accuracy deteriorates when there is a difference between the actual height and assumed height of the road. In order to solve this problem, the authors proposed a three dimensional vehicle tracking method that combines the elevation and azimuth measured by an image sensor, and range and azimuth measured by a two dimensional radar. In this method, it is assumed that the road surface is flat; the first two samplings only obtain the initial state of the vehicle position and velocity, because three dimensional position is not measured by image and two dimensional radar, directly. After obtaining an initial state, the vehicle state is updated by measurement from both sensors by using a linear approximation of their measurement models based on an extend Kalman filter technique.
Samenvatting