| Popis: |
In autonomous driving vehicles, recognizing objects in the environment, such as other vehicles, pedestrians, crossroads, lanes, and curbs, is necessary. Curbs separate roads from sidewalks, and if they are properly detected and recognized, autonomous vehicles would be prevented from accidentally encroaching onto sidewalks. Various distance-measuring sensors such as radars, lidars, stereovision sensors, and ultrasonic sensors can be used to obtain a vehicle’s distance from the curb. However, a cost-effective curb-detection and recognition system that is robust under various weather and light conditions is desirable for most vehicles. Automotive ultrasonic sensors are good candidates for this application owing to their low cost, and as they are already widely used in most vehicles—especially for parking assistance. Although ultrasonic sensors are useful for measuring distances to curbs, their observed performance during our field tests was poor, with frequent outliers and unreliable distance outputs. We thus attempted to overcome these limitations by using multiple ultrasonic sensors simultaneously. We tested an averaging algorithm and a majority-voting algorithm as distance-estimation algorithms using measurements from three ultrasonic sensors mounted on a vehicle. The distance-estimation performance obtained was improved but still insufficient for reliable curb detection. We found that the measurements were sometimes being made from the ground instead of the curb, which significantly degraded the distance-estimation performance. The ultrasonic wave transmitted from the sensor can be reflected from any object within the beam width of the antenna attached to the sensor, even when the antenna is properly oriented toward the curb. We therefore present ground-reflection elimination algorithms for ultrasonic sensors for enhanced distance-estimation performance and verify their effectiveness through field tests. Three ultrasonic sensors were installed on the lower side of a vehicle for the experiment. We compared the distance-estimation performance with and without the ground-reflection elimination and demonstrated significant performance enhancement with the proposed algorithms. |
