[Paper Review] High-Definition Map Generation Technologies For Autonomous Driving

Zhibin Bao, Md. Sabir Hossain+2 more|arXiv (Cornell University)|Jun 11, 2022

Robotics and Sensor-Based Localization18 citations

TL;DR

This paper reviews state-of-the-art high-definition (HD) map generation technologies for autonomous driving, focusing on 2D and 3D map synthesis using multi-sensor data (LiDAR, cameras, IMU, GPS) and advanced algorithms. It synthesizes current methodologies, identifies key limitations, and outlines future research directions to improve map accuracy, scalability, and real-time applicability in autonomous systems.

ABSTRACT

Autonomous driving has been among the most popular and challenging topics in the past few years. On the road to achieving full autonomy, researchers have utilized various sensors, such as LiDAR, camera, Inertial Measurement Unit (IMU), and GPS, and developed intelligent algorithms for autonomous driving applications such as object detection, object segmentation, obstacle avoidance, and path planning. High-definition (HD) maps have drawn lots of attention in recent years. Because of the high precision and informative level of HD maps in localization, it has immediately become one of the critical components of autonomous driving. From big organizations like Baidu Apollo, NVIDIA, and TomTom to individual researchers, researchers have created HD maps for different scenes and purposes for autonomous driving. It is necessary to review the state-of-the-art methods for HD map generation. This paper reviews recent HD map generation technologies that leverage both 2D and 3D map generation. This review introduces the concept of HD maps and their usefulness in autonomous driving and gives a detailed overview of HD map generation techniques. We will also discuss the limitations of the current HD map generation technologies to motivate future research.

Motivation & Objective

• To provide a comprehensive review of recent advancements in high-definition (HD) map generation for autonomous driving.
• To analyze the integration of multi-sensor data (LiDAR, cameras, IMU, GPS) in HD map construction.
• To identify technical limitations in current HD map generation methods that hinder scalability and real-time deployment.
• To guide future research by highlighting gaps in accuracy, generalization, and computational efficiency.

Proposed method

• Systematic review of 2D and 3D HD map generation techniques using multi-modal sensor data.
• Categorization of methods based on data sources, fusion strategies, and representation formats (e.g., vector maps, semantic segmentation).
• Analysis of deep learning and geometric modeling techniques used in feature extraction and map reconstruction.
• Evaluation of localization accuracy and map update frequency in dynamic environments.
• Comparison of manual vs. automated map creation pipelines in terms of cost, precision, and scalability.
• Identification of key challenges such as data sparsity, occlusion handling, and real-time processing constraints.

Experimental results

Research questions

• RQ1 What are the dominant sensor fusion and mapping techniques used in modern HD map generation for autonomous vehicles?
• RQ2 How do current HD map generation methods balance accuracy, computational efficiency, and real-time performance?
• RQ3 What are the primary limitations in existing HD map technologies that hinder full deployment in dynamic urban environments?
• RQ4 How do 2D and 3D map representations compare in terms of localization precision and semantic richness?
• RQ5 What future research directions are most critical to overcome current bottlenecks in HD map scalability and automation?

Key findings

• HD maps significantly enhance localization accuracy in autonomous driving by providing centimeter-level precision.
• Multi-sensor fusion, especially combining LiDAR and camera data, improves map robustness and semantic detail.
• Current methods face challenges in handling dynamic objects and occlusions, leading to map inconsistencies.
• Automated map generation remains limited by high computational costs and reliance on high-quality training data.
• Real-time HD map updates are still not reliably achieved in large-scale urban environments.
• There is a growing trend toward learning-based, end-to-end map generation, but generalization across diverse environments remains a key open challenge.