Segmented convolutional gated recurrent neural networks for human activity recognition in ultra-wideband radar

Abstract

The automatic detection and recognition of human activities are valuable for physical security, gaming, and intelligent interface. Compared to an optical recognition system, radar is more robust to variations in lighting conditions and occlusions. The centimeter-wave ultra-wideband radar can even track human motion when the target is fully occluded from it. In this work, we propose a neural network architecture, namely segmented convolutional gated recurrent neural network (SCGRNN), to recognize human activities based on micro-Doppler spectrograms measured by the ultra-wideband radar. Unlike most existing approaches which treat the micro-Doppler spectrograms the same way as natural images, we extract segmented features of spectrograms via convolution operation and encode the feature maps along the time axis with gated recurrent units. Taking advantage of regularities in both the time and Doppler frequency domains in this way, our model can detect activities with arbitrary lengths. The experiments show that our method outperforms existing models in fine temporal resolution, noise robustness, and generalization performance. The radar system can thus recognize human behavior when visible light is blocked by opaque objects.