Core Concepts
The author proposes a method to automatically annotate pole-like features in images using 2D HD maps and lidar sensors, demonstrating the effectiveness of object detection for pole base detection.
Abstract
The content discusses the use of high-definition maps (HD maps) for automatic annotation of pole-like features in images. It introduces a framework that leverages lidar sensors to refine annotations by filtering out occluded features and refining ground projections. The methodology is validated using data from semantic segmentation datasets and real-world data from the city of Compiègne, France. The study highlights the importance of annotated training data for deep learning methods in image-based approaches.
The paper outlines two types of HD maps: low-level features like point clouds or image-based key points/frames, and vector maps encoding road infrastructure at a higher semantic level. It emphasizes the significance of georeferenced features as landmarks for vehicle pose estimation. The approach involves projecting map features onto image frames, defining pole bases at ground level, and utilizing object detection techniques for detection.
Furthermore, the study delves into lidar-based refinement and filtering processes to enhance annotation accuracy. It details how lidar sensors are used to estimate ground surfaces, filter occluded features, and refine height estimations. The experimental results showcase the performance metrics obtained through training models on different datasets with varying box sizes.
In conclusion, the research presents a promising framework for automating pole base detection using HD maps and lidar sensors. It suggests avenues for future improvements such as tuning parameters for better filtering, expanding training data variability, and exploring applications beyond pole base detection.
Stats
"A map feature is therefore represented as a 2D point MP with coordinates Mx, My expressed in the map frame."
"We consider only vector maps as they are sensor agnostic."
"Using all true positive predicted bounding boxes, we computed the Mean Absolute Error (MAE) between predicted x-coordinate and labeled one."
Quotes
"No height information available; it is not possible to project map features directly onto an image."
"Annotating images using HD maps with lidar-based filtering improves annotations but may introduce some noise."
"The automated annotation framework enables easy addition of new training data without additional labeling costs."