Object Detection in EO/IR and SAR Images Using Low-SWAP Hardware

This paper studies the performance of object detection algorithms applied to electro-optic (EO), infrared (IR), and synthetic aperture radar (SAR) data. First, we describe the simulation of EO and IR images containing vehicles and people and the merging of measured environmental SAR scenes with target image chips. This produces a large data set for training three deep learning algorithms: RetinaNet, EfficientDet, and YOLOv5. The algorithms were trained with a powerful elastic compute cloud (EC2) instance. Performance on simulated data at inference time, in terms of speed and accuracy, was tested on the EC2 instance and a low size weight and power (SWAP) single board computer. YOLOv5 was the most accurate algorithm and the fastest on the EC2 instance but the slowest on the low-SWAP device. RetinaNet and EfficientDet and produced operationally useful throughput on the low-SWAP device for surveillance applications, with RetinaNet having the higher accuracy. This is believed to be the first time the same algorithms have simultaneously been tested on EO, IR, and SAR data, and compared in a unified framework.