Human Detection by Deep Neural Networks Recognizing Micro-Doppler Signals of Radar

The purpose of this paper is to show the effectiveness of Deep neural networks (DNN) for recognizing the micro-Doppler radar signals generated by human walking and background noises. To show this, we collected various micro-Doppler signals considering the actual human walking motion and background noise characteristics. Unlike the previous researches that required a complex feature extraction process, we directly use the FFT result of the input signal as a feature vector without any additional pre-processing. This technique helps not to use heuristic approaches to get a meaningful feature vector. We designed two types of DNN classifier. The first is the binary classifier to classify human walking Doppler signals and background noises. The second is the multiclass classifier that is roughly able to recognize a circumstance of a place as well as human walking Doppler signals. DNN for the binary classifier showed about 97.5% classification accuracy for the test dataset and DNN(ReLU) for the multiclass classifier showed about 95.6% accuracy.