The Lifelines of Autonomous Mobility: 3D Road Networks

Written by: Ibrahim Shoer

Just as blood vessels are essential for transporting life-sustaining nutrients throughout the body, roads are the lifeblood of our communities, facilitating movement, trade, and connectivity. In the realm of autonomous driving, the intricate web of streets, highways, and alleys becomes even more critical. These networks are not just passive surfaces for vehicles to traverse; they are dynamic, interactive ecosystems that need to be deeply understood and modeled to ensure the safe, efficient flow of future traffic.

Importance of Modeling 3D Road Networks

The leap from human-operated vehicles to autonomous driving systems is monumental. Autonomous vehicles (AVs) require an intricate understanding of their environment to navigate safely and efficiently. Here, 3D road network modeling becomes indispensable. It provides a comprehensive framework that captures the complex interplay of various road elements—elevations, lanes, intersections, and traffic controls. 3D based tools offer realistic 3D road network-based simulation environments, providing a viable alternative to real-world testing. These simulations are crucial for assessing vehicle performance and improving algorithms without the risks and costs associated with physical testing.

Diversity of Road Topologies

Just as no two cities are alike, their road networks reflect their unique histories, geographies, and cultures. Some cities boast grid-like patterns, facilitating straightforward navigation but potentially leading to congestion. Others feature intricate, organic layouts that might challenge navigation yet offer multiple paths to any given destination. Each topology comes with its advantages and challenges for autonomous driving:

• Grids (e.g., New York City): Predictable and easy to navigate, but can suffer from the "canyon effect" leading to GPS inaccuracies.
• Radial (e.g., Paris): Efficient for distributing traffic from the center, beneficial in wartime for defense and control, but can be confusing for autonomous systems due to the many angled intersections and roundabouts.

The radial design of Paris, with avenues radiating from landmarks such as the Arc de Triomphe, was crucial during wartime. Originating with strategic defenses like the Thiers wall, Paris's layout has developed over time, showcasing the city's ability to adapt and evolve in its approach to urban design.

• Organic (e.g., Boston): Reflects natural terrain and historical growth, offering multiple routes but presenting significant challenges in mapping and predictability.

Modeling of City Road Networks

The goal isn't just to understand existing road networks but to leverage this understanding to create enhanced 3D models of cityscapes. By analyzing cities like Paris or Rome, we can extract the unique topological characteristics that define them. Roads can be approximated to a graph structure, where intersections become nodes, and streets become connecting edges. Utilizing Graph Neural Networks (GNNs), these models can learn the deep features of different topologies, allowing for the generation of new, optimized 3D cities. This approach enables the crafting of road networks that maintain the topological and functional essence of any city while enhancing efficiency and safety for autonomous vehicles.

Following our goal of synthesizing 3D road networks, we had to study the topologies and possible means to generate new networks.

In our study, road network data from roughly 40,000 cities was harnessed from OpenStreetMap and restructured into graph network form.?

Graph Neural Networks (GNNs) were utilized to derive embeddings from this real-world data, capturing the essence of structural road characteristics, This is inspired by the concept of graph-graph proximity, where we utilize Graph Convolutional Networks (GCNs) to learn embeddings for pairs of graphs.

To achieve this, we use the Laplacian matrix distance between graph pairs as the ground truth during training. The Laplacian matrix provides a fundamental measure of graph connectivity and structure, enabling us to establish a meaningful measure of similarity between different graphs.

Well, embeddings are useless unless they offer insights. So embeddings are segmented into clusters via K-means clustering, an algorithm to group together similar road networks to uncover inherent patterns in the embeddings, which signify key structural similarities among the road networks. This analysis helps categorize clusters into distinct bands of low, moderate, or high network properties, enhancing our understanding of the underlying network structures.

In the final stage of our work, we focus on the generation of synthetic road networks using GAE, leveraging the tags generated from our cluster analysis. This user-centric approach allows for the generation of road networks with specific attributes such as ”high density” or ”low node counts” or even similar to “Paris”.

Take a look at these examples from Cluster 1 and Cluster 2 of synthesized road networks:

Notice the distinct density differences between the two clusters. Now, consider the possibilities if you had access to a tool that could create a wide array of road networks. What innovative ideas or applications would you come up with? Share your thoughts and ideas in the comments section below! ??

Armed with this tool and our procedural models, possibilities abound. Here are just a few applications:

• Recreate your favorite city, customizing districts and altering roads to alleviate traffic congestion, or strategize reconstruction for rush hour scenarios.
• Design new road networks for game environments, fostering community development within them.
• Create diverse testing grounds for autonomous vehicles, incorporating varied road features, vegetation, and even buildings.

Look out for the comprehensive article in the upcoming #ICMLA proceedings for more insights.