Multi‐lane recognition using the YOLO network with rotatable bounding boxes

Abstract Currently, driver assistance and autonomous driving functions are emerging as essential convenience functions in automobiles. For autonomous driving, fast and accurate lane recognition is required, along with driving environment recognition. The recognized lanes must be divided into ego and left‐ and right‐side lanes. Among deep learning, the You Only Look Once (YOLO) network is widely known as a fast and accurate object detection technique. The general methods are not robust to angle variations of the objects because of the use of a traditional bounding box, a rotation variant structure for locating rotated objects. The rotatable bounding box (RBBox) can effectively handle situations where the orientation angles of the objects are arbitrary. This study uses a YOLO approach with RBBox to recognize multi‐lane accurately. The proposed method recognizes the ego lane and its surrounding lanes by accurately distinguishing them. And the proposed method shows stable multi‐lane recognition performance by predicting them that exist in the images but do not exist in the ground truth of the TuSimple data set. Even compared to other lane recognition methods, it shows good competitiveness. Nevertheless, more training data and network learning are needed in a specific road environment (a lane is centered on the image).