Single-camera based vehicle pose estimation using multiple features on the road surface

Abstract

For autonomous navigation, it is essential to estimate the pose of vehicles. Many existing pose estimation methods are based on global positioning system (GPS) and inertial motion unit (IMU), which show some limited performance due to GPS shadow area and IMU error accumulation. To eliminate the limitations, the previous works estimate the vehicle pose by vision sensors. However, the vision-based pose estimation methods assume that there are enough features to be matched, thus show limitations when the environment consists of featureless regions. Especially, the multi-camera frameworks such as stereo, do not work well, because there are only a few matched features among the multi-camera images. For the general traffic scenes with many moving objects, the vision-based pose estimation methods can be biased due to the image features of moving objects. In addition, when the camera has only a forward-motion or a relatively small base line, the triangulation uncertainty of 2-D feature matching becomes large and the pose estimation errors increase. For solving the limitations, this thesis presents a hierarchical feature network and a pose estimation method without triangulation. The hierarchical feature network consists of road regions, road lanes, and road point features. The road lanes and the road point features are involved in the road regions, so they can be extracted from the road regions or used as important cues. In this thesis, the vehicle pose is estimated by the relations among the features. The vehicle pose estimation is performed using the matches of the road point features, which are extracted from the road regions, and the road regions are segmented from the road lane features. Assuming the road features static, the proposed features are better suited to the vehicle pose estimation on general traffic scenes, which involve many moving objects. Additionally, the proposed algorithm estimates the vehicle pose using the optimal 3-D points, which are extrac...