Optimization and Deployment of Real-Time On-Orbit Intelligent Interpretation Algorithms for Spaceborne Remote Sensing

Area of Science:

• Computer Vision
• Aerospace Engineering
• Embedded Systems

Background:

• Orbital remote sensing demands real-time object detection for situational awareness.
• Deploying Convolutional Neural Network (CNN)-based models on spaceborne edge devices is hindered by Size, Weight, and Power (SWaP) constraints.
• Conventional single-stage detectors' branch-and-merge topology complicates Field-Programmable Gate Array (FPGA) implementation due to memory usage and pipeline stalls.

Purpose of the Study:

• To propose RS-YOLO, an object detection algorithm addressing SWaP limitations for spaceborne edge computing.
• To enhance the efficiency of CNN-based object detection on resource-constrained platforms.
• To achieve a balance between hardware efficiency and detection performance in orbital remote sensing.

Main Methods:

• Developed RS-YOLO using a hardware-software co-design approach.
• Implemented structural re-parameterization to create a sequential stream of padding-free convolutions, optimizing dataflow.
• Introduced a spatial heterogeneous quantization (SHQ) engine to mitigate quantization errors for small targets.

Main Results:

• The RS-YOLO system achieved a mean Average Precision (mAP) of 0.887 on the NWPU VHR-10 dataset.
• The system demonstrated low power consumption (2.24 W) on the Xilinx Zynq-7020 platform.
• Achieved high energy efficiency of 15.19 GOPS/W, with only a 1.4% mAP decrease compared to FP32 baseline.

Conclusions:

• RS-YOLO effectively balances hardware efficiency and detection performance for resource-constrained edge platforms like micro-satellite payloads.
• The proposed hardware-software co-design and SHQ engine offer a viable solution for real-time object detection in spaceborne applications.
• Demonstrated the feasibility of deploying advanced CNN-based object detection on edge devices with stringent SWaP limitations.