3D modeling for the entire region of Seoul with 25,000 aerial photographs
The above video shows the outcome of 3D modeling for the entire region of Seoul–spanning over 605 km2 with over 600,000 buildings–created by NAVER LABS. It may appear as a video filmed by a drone, but in fact, it shows 3D models covered with textures to make them look realistic.
NAVER LABS, which has been working with the Seoul Metropolitan Government to build the Virtual Seoul platform, has the technology to capture urban-based 3D models with aerial photographs. This technology restores 2D images of large areas taken from the sky into highly precise 3D models and naturally connects them to each other. These 3D models of Seoul were created in just 30 days processing approximately 25,000 aerial photographs taken over the course of 17 days.
This 3D spatial data is useful for many purposes; the data allows the Seoul Metropolitan Government to analyze virtualized real-world spaces and utilize them for monitoring or simulation to tackle urban development projects or environmental issues. NAVER LABS also uses these 3D models to create high-precision maps, which serve as the key data for autonomous driving vehicles.
AI technology and aerial photo processing capabilities of NAVER LABS
The conversion of aerial photographs into 3D requires a technique called photogrammetry. How can we extract 3D information from 2D photographs? The answer lies in leveraging the disparity between different photographs. If we look at the photographs taken on the move, the closer the location is to the camera, the greater the change in its position, and the farther it is, the smaller the change. Analyzing this change also allows us to distinguish between the building roof and the ground and to calculate its height from aerial photographs.
Certainly, the technologies to accurately match and link each data are also important as we utilize thousands of photographs. The know-how and solutions to this are also vital to urban-sized 3D modeling.
1) Image matching : Connects the same points on thousands of photographs
2) Bundle adjustment : Estimates the correct poses of photographs using ground control points
3) Dense matching : Applies algorithms to create 3D models for positions with no or insufficient feature points and those with frequent matching errors
4) 3D modeling : Aggregates estimated dense matching results from many images obtained by shooting the same region from different locations to create a continuous 3D digital surface
Many of the spaces we live in continue to undergo digitalization. New data leads to new solutions. The urban-sized 3D modeling technology will also allow us to understand the vast everyday space of cities from a new perspective, while at the same time, being used for building future infrastructures such as autonomous driving.
NAVER LABS will be updating its 3D models for Seoul and HD maps based on new photographs taken in 2020.
A more detailed description of this technology is available in the article below.
Aerial image based 3D modeling for HD mapping