RarePlanes Detection Using YOLOv5 on the Versal Adaptive SoC

Object detection is an important operation for many satellites performing remote sensing operations. Machine learning models are increasingly used for real-time object detection. It is crucial that such machine learning models perform inference on satellite images accurately and reliably, while meeting power, size, and other constraints. We train several YOLOv5 object detection models on the RarePlanes dataset, a collection of satellite aerial images of airplanes, and deploy the trained models to the Deep Learning Processing Unit on the Versal AI Core XCVC1902 device. The performance and accuracy of the models running on the Versal device are measured and compared to implementations on a CPU and GPU. Radiation-induced single-event upsets are then simulated by injecting faults into the programmable logic of the Versal device, and the effects of each fault on the YOLOv5 model are discovered.