Vehicle Perception From Satellite.

Area of Science:

• Computer Vision
• Remote Sensing
• Urban Mobility

Background:

• Satellites offer a unique, broad city-scale view for traffic monitoring, surpassing traditional methods.
• Challenges include tiny vehicle size, slow movement, and dynamic backgrounds due to Low Earth Orbit (LEO) satellites.
• Existing computer vision methods face difficulties with these specific conditions.

Purpose of the Study:

• To establish a large-scale benchmark dataset for satellite-based traffic monitoring.
• To support research in critical tasks like tiny object detection, vehicle counting, and density estimation.
• To encourage further research and development in this emerging field.

Main Methods:

• Dataset construction using 12 real satellite videos and 14 synthetic videos from GTA-V.
• Creation of 408 video clips with 7,336 real and 1,960 synthetic images.
• Annotation of 128,801 vehicles across all images, with varying vehicle counts per image.

Main Results:

• The benchmark dataset facilitates evaluation of various computer vision approaches for satellite traffic monitoring.
• Analysis of performance differences between classic and state-of-the-art methods is provided.
• Identified key challenges and discussed future research directions for satellite-based traffic analysis.

Conclusions:

• The developed benchmark is crucial for advancing vehicle perception and traffic monitoring from satellite imagery.
• It provides a standardized platform for evaluating algorithms under realistic, challenging conditions.
• The dataset and analysis pave the way for improved traffic management applications using remote sensing data.