Fast and Accurate System for Onboard Target Recognition on Raw SAR Echo Data

Synthetic Aperture Radar provides high-resolution imaging independent of 1 weather conditions. SAR data is acquired by an aircraft or satellite and sent to a ground 2 station to be processed. For novel applications requiring on-time processing for real-time 3 analysis and decisions, onboard processing is necessary to avoid the limited down-link ca- 4 pacity and reduce data analysis latency for real-time decisions. Real-time target recognition 5 from SAR data has emerged as an important task in many areas, like defense and surveil- 6 lance where targets must be identified as fast as possible. The accuracy of target recognition 7 algorithms has improved by applying deep learning models. These are computationally 8 and memory expensive, which requires optimized models and architectures for efficient 9 deployment in onboard computers. This paper presents a fast and accurate target recogni- 10 tion system directly on raw SAR data. The neural network model receives and processes 11 SAR echo data for fast processing, without computationally expensive DSP image genera- 12 tion algorithms such as BackProjection and RangeDoppler. This permits the utilization of 13 simpler and faster models, but still very accurate. The system was designed, optimized, 14 and tested on low-cost embedded devices with low energy requirements, namely a Khadas 15 VIM3 and a Raspberry Pi 5.The solution achieves a target classification accuracy for the 16 Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset close to 100% 17 in less than 1.5 ms and 5.5W of power.