A Large-Scale Synthetic Benchmark Dataset for Non-Cooperative Space Target Perception.

Area of Science:

• Aerospace Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Accurate space target perception is vital for on-orbit aerospace missions.
• Deep learning models show promise for space target perception but require large labeled datasets.
• Existing datasets are insufficient for comprehensive deep learning model training.

Purpose of the Study:

• To address the limitations of current datasets for space target perception.
• To build a multitask synthetic benchmark dataset named NCSTP.
• To support simultaneous space target detection, recognition, and component segmentation.

Main Methods:

• Collected and modified models of satellites, space debris, and space rocks.
• Generated 200,000 synthetic images in a realistic space environment using Blender.
• Annotated data for detection, recognition, and component segmentation tasks.

Main Results:

• Developed the NCSTP dataset with diverse space target variations.
• The dataset supports multitask learning for perception tasks.
• Established a benchmark by testing state-of-the-art models on the dataset.

Conclusions:

• The NCSTP dataset effectively addresses the need for large-scale labeled data in space target perception.
• The benchmark facilitates the evaluation and advancement of deep learning models for aerospace applications.
• This resource will accelerate the development of autonomous perception systems for space missions.