Abstract
Research and/or Engineering Questions/Objective
With the recent growth of interest in autonomous vehicle, estimating vehicle location has become one of the core technologies in the automotive field. However, low-cost global positioning systems (GPS) has too significant errors to meet accuracy requirements of autonomous driving. To reduce such errors, the map-matching method has been developed. The map-matching algorithm compensates GPS positioning errors by matching extracted features from sensors with the database map. The road surface marker (RSM) is one of the features used for map-matching. The objective of this paper is to extract RSM precisely in various conditions to enhance the precision of localization.
Methodology
The vision-based RSM detection algorithm is affected by surrounding environments of the ego vehicle. In particular, shadows and surrounding objects are critical to the algorithm output because they are hard to distinguish from the RSM by camera. In response, this paper suggests two methodologies to enhance the robustness of RSM detection. The first methodology is to eliminate shadows on images, and the second is to distinguish surrounding objects by using a Light Detection And Ranging (LIDAR) sensor. First, LIDAR detects the location of objects around the ego vehicle. Next, through coordinate conversion of LIDAR data, the objects on image are excluded from the region of interests (ROI). Then, the shadow elimination algorithm and the filtering algorithm are applied on the ROI of the image. Finally, adaptive thresholding is performed to achieve a result image. The result image is applied to the map-matching algorithm to compensate for the GPS positioning error.
Results
We evaluated the proposed algorithm in several types of challenging scenes in various weather conditions. The first type displays low contrast between the road and the RSM. The second type has shadows that are difficult to distinguish from the RSM. The third type has different intensity distributions between areas that make it difficult to extract RSM by a single threshold value. The fourth type shows crowded surrounding vehicles. The performance is graded based on location errors between the estimated vehicle position and the reference vehicle position. The position from Real-Time Kinematics GPS (RTK GPS) is used as the reference.
Limitations of this study
The precision of vehicle localization based on the map-matching algorithm is affected by the accuracy of the map. This study has shown limitation based on the fact that the accuracy of the map was not taken into consideration. Although the reference value from the RTK GPS has a higher level of precision than that of low-cost GPS, it also contains errors. Therefore, the reference value may not be the exact true value.
What does the paper offer that is new in the field including in comparison to other work by the authors?
The RSM extraction algorithm in this paper contains a shadow removal algorithm. Additionally, LIDAR data is used to eliminate surrounding obstacles in image planes. Furthermore, the evaluation parameter of this paper is localization precision. These three key points are new in comparison to other works done in relevant fields.
Conclusions
To enhance the precision of map-matching-based localization, the accuracy of extracted features must be guaranteed. Accordingly, this paper proposes a vision-based robust RSM extractor to ensure precise localization.
KEYWORDS – road surface marker (RSM), vehicle localization, lane detection, computer vision