Automatic Point Cloud Annotation using existing HD Map Data for Map Construction

Kurzfassung

Perception is a critical task for autonomous driving and is challenging because of the scale and complexity of urban environments, which leads to limitations in sensor accuracy. Point-wise segmentation plays a crucial role in addressing these challenges. However, point cloud datasets have limitations related to data storage capacity, dimensionality, and the scale factor during manual annotation, making it difficult to achieve efficiency in terms of time and cost. In this study, we propose an automatic annotation pipeline sourced from another key contributor to automated driving systems: high-definition (HD) maps. HD maps provide lane-level information regarding traffic scenarios, 3-dimensional coordinates, and semantic classes of objects in a lightweight format. Our approach relies on generating a synthetic point cloud from these specific vector-type datasets and matching it with the original sensor-collected training dataset. We then utilize the point clouds and this information to train and generate reliable predictions. Ultimately, we propose that this preprocessing stage can serve as a baseline for producing new HD maps, thereby contributing to automated driving in real-world applications.