A Lightweight DCNN Algorithm for Radar-Based Suspicious Human Activities Classification With Data Augmentation Techniques

Recent research on deep convolutional neural networks (DCNNs) for classifying micro-Doppler (m-D) signatures faces a significant challenge due to the limited and costly acquisition of data. This constraint restricts the depth and performance of DCNN models that can be effectively implemented. Moreover, most DCNN models are inherently complex and computationally expensive, which exacerbates issues of overfitting in scenarios with limited data availability. Using pre-trained models poses disadvantages when applied to small-scale datasets of m-D signatures. Pre-trained models, trained on unrelated data such as images, often fail to capture the unique characteristics of radar m-D signatures, resulting in suboptimal performance. In this study, we address these challenges by introducing augmentation methods tailored for radar m-D signatures. Techniques such as time shift, frequency disturbance, and frequency shift are utilized to generate diverse datasets, thereby enriching the training data and enhancing model generalization. Furthermore, we propose a lightweight DIAT-DSCNN-HARNet model specifically designed for this task. Our experiments demonstrate that this model outperforms both pre-trained and scratch DCNN models in terms of classification accuracy for augmented data. Specifically, our approach shows superior performance compared to images generated by Generative Adversarial Networks (GANs). Our findings highlight the effectiveness of augmentation strategies and specialized model design in improving m-D signature classification.