Human Activity Recognition Based on Deep Learning Method

Xiaoran Shi,Yaxin Li,Feng Zhou,Lei Liu

DOI: https://doi.org/10.1109/radar.2018.8557335

2018-01-01

Abstract:With the increasing demand of security defense, anti-terrorism investigation and disaster rescue, human activity classification and recognition have become a hot research topic. When a human is illuminated by electromagnetic waves, a Doppler signal is generated from his or her moving parts. Indeed, bodily movements are what make humans' micro-Doppler signatures unique, offering a chance to classify human activities. Classification needs a lot of samples for training, however, in the real application, there is a certain gap between the simulated data and the real data, and the measured data is often difficult to obtain. Due to the non-stationary characteristic for human radar echoes, the spectrograms for the human activities show different micro-Doppler signatures. Therefore, we proposed a method of human activity classification based on spectrograms using deep learning techniques, including deep convolutional generative adversarial network for expanding and enriching training set and a transfer-learned deep convolutional network (DCNN) for feature extraction and classification, which is based on a DCNN pre-trained by a large-scale RGB image data set-that is, ImageNet. Finally, the simulation results verified the effectiveness of the proposed method.