Wall Effect Mitigation for Through-the-Wall Human Motion Detection Using a GAN Network

In the context of through-the-wall human motion detection, the DC clutters, indoor multipath clutters and environmental noises introduced by the wall media and indoor static objects often obscure the important motion information, thus affecting the effective recognition of the target motions. Many strategies have been proposed to suppress such effect, including spatial filtering, high-pass filtering, subspace decomposition, etc. However, these methods either rely heavily on manual intervention or not robust to noises. This paper attempts to cope with these challenges using Generative Adversarial Networks (GAN). More specifically, a Pix2Pix network is applied to learn the mapping from the wall corrupted raw range map to its de-walled counterpart. Our implement shows that the DC clutters and noises can be effectively removed in the raw range map, with human motion information being preserved with high fidelity. Also, this result outperforms our previous proposed optimization based RPCA approach.