Transfer Learning for Human Activities Classification Using Micro-Doppler Spectrograms

Abstract

Human activities classification has drawn great attention due to its potential applications in security, surveillance and gesture-based interface. The movements of the human body and limbs result in unique micro-Doppler features which can be exploited for identification of human behavior. In this work, we propose a transfer-learned residual network to classify human activities based on micro-Doppler spectrograms. The residual network (ResNet) is pre-trained on ImageNet and fine-tuned on an empirical non-parametric human model using Motion Capture Database. Compared with typical CNN from scratch, this ResNet-based method requires shorter epochs (within 50 epochs) and achieves higher accuracy (rise nearly 6% on the average classification accuracy) for micro-Doppler spectrograms classification. Apart from statistical evaluation, we implement guided backpropagation method and t-Distributed Stochastic Neighbor Embedding (t-SNE) technique to visualize the transfer learning of residual network using spectrograms.