armasuisse S+T – Edge AI for Satellites and Drones

Challenge

Satellite images are traditionally sent to the ground for analysis. For time-critical applications like disaster detection, this is too slow. Image analysis must happen directly on the satellite or drone.

Solution

• Phase 1: Image processing pipeline with TensorFlow for calibration and geocorrection
• Phase 2: YOLO-based object detection on FPGA hardware
• Phase 3: End-to-end demo on drone with real-time analysis

Our Contribution

• Development of the complete image processing pipeline
• Model training and optimization for FPGA deployment
• Integration and testing on Xilinx FPGA hardware
• Building the end-to-end demonstrator system

Technologies

Results

Functioning demonstrator for edge computing: 10.5 megapixels per second inference performance at only ~26 watts power consumption. The solution is deployable on satellites, drones, or autonomous vehicles, demonstrating that complex image analysis is possible even under extreme resource constraints.

More about our space projects at ateleris.space