Automatic Target Recognition (ATR) is one of the strongest growing technology topics, especially in commercial applications like image recognition and automotive. ATR applications in radars established in the fields of defense and protection stayed on a certain level for years, so today there is a strong need to participate in the performance growth enabled by modern technologies. In this paper the ATR potential of the Deep Learning (DL) approach is proved by comparing the performance of a highly mature “classical” concept using feature extraction and several types of classifiers versus several state of the art DL networks and training concepts. The comparison is done on the same large database of Micro Doppler signatures from several classes, as typically used for moving object classification and drone threat assessments. The comparison shows the benefit and potential drawbacks of the different approaches.
Paper
Full text
Deep Learning Approach for Radar Applications
Semantic Scholar · Engineering · 2020
Abstract
Automatic Target Recognition (ATR) is one of the strongest growing technology topics, especially in commercial applications like image recognition and automotive. ATR applications in radars established in the fields of defense and protection stayed on a certain level for years, so today there is a strong need to participate in the performance growth enabled by modern technologies. In this paper the ATR potential of the Deep Learning (DL) approach is proved by comparing the performance of a highly mature “classical” concept using feature extraction and several types of classifiers versus several state of the art DL networks and training concepts. The comparison is done on the same large database of Micro Doppler signatures from several classes, as typically used for moving object classification and drone threat assessments. The comparison shows the benefit and potential drawbacks of the different approaches.