Application of DNN for radar micro-doppler signature-based human suspicious activity recognition

Abstract

The non-availability of open-source datasets covering various human suspicious activities and well-trained deep learning (DL) architecture limits the effective utilization of DL network-supported radar systems for real-time autonomous human activity recognition (HAR). The development of a dataset and validation of its micro-Doppler-signature-distinguishable-features by a suitable DL network becomes significant, and that is the key contribution given in this research. In this work, an indigenously developed X-band CW radar is employed to create a diverse DIAT-μ RadHAR dataset, which includes (a) army marching, (b) Stone pelting/Grenades throwing, (c) jumping with holding a gun, (d) army Jogging, (e) army crawling and (f) boxing activities. Six Pre-trained CNNs models supported DL architectures, trained with DIAT-μRadHAR dataset containing 3780 m-D images, are proposed, which can be used for open-field HAR. The characteristics of our dataset and performance analysis of the proposed two DL architectures are statistically computed, and their results are compared in terms of receiver operating characteristic (ROC), precision, F1-score, recall, and confusion matrix. The pre-trained VGG19 with transfer learning outperforms other CNNs with an overall classification accuracy of 98%.