少数の学習データセットによる単眼カメラを用いた道路標示検出精度の向上策の検討

抄録

Research on automated driving is currently being conducted by various manufacturers around the world. Many Japanese manufacturers are using HD (High Definition) map, which has the information on road features with position. The automated driving vehicle will comparing several types of road features in the HD map and features detected by the sensors on vehicle during driving information on road markings, road signs, and lane locations.

In this study, we investigated a method to improve detection accuracy using a limited data set by adding a process to suppress false positives to machine learning framework, such as YOLO, an existing object detection network, for 15 types of labels using images captured by monocular camera. With a set of experimental data, we confirmed a 40% improvement in the accuracy rate compared to the case where only object detection is performed with a low detection threshold.