Localization and classification of space objects using EfficientDet detector for space situational awareness.

Area of Science:

• Space Situational Awareness (SSA)
• Computer Vision
• Deep Learning

Background:

• Space situational awareness (SSA) is crucial for space navigation, requiring accurate recognition of space objects like spacecraft and debris.
• Challenges in space object recognition include varied object sizes, high contrast, low signal-to-noise ratio, and noisy backgrounds.
• Existing convolutional neural network methods often struggle with accuracy due to attention loss on objects in complex space imagery.

Purpose of the Study:

• To propose a robust decision fusion method for improved space object detection and classification.
• To address the limitations of existing methods in handling complex sensing conditions in space.
• To enhance the accuracy and reliability of space object recognition for SSA systems.

Main Methods:

• A decision fusion method was developed, training an EfficientDet model with an EfficientNet-v2 backbone for initial object detection.
• Detected objects were augmented with blurring and noise before being processed by an EfficientNet-B4 model.
• Decisions from both models were fused to classify objects into 11 categories using the SPARK dataset.

Main Results:

• The proposed method achieved superior performance in object detection and classification.
• Demonstrated significant improvements in accuracy (94%) and performance metric (1.9223%) for classification.
• Achieved high mean precision (78.45%) and mean recall (92.00%) for object detection.

Conclusions:

• The decision fusion approach significantly enhances space object recognition capabilities.
• The method is feasible for real-world space situational awareness systems.
• The study highlights the effectiveness of combining multiple deep learning models for complex visual recognition tasks in space.