CAE-MAS: Convolutional Autoencoder Interference Cancellation for Multiperson Activity Sensing With FMCW Microwave Radar
Hossein Raeis, Mohammad Kazemi, Shervin Shirmohammadi
Abstract
Human Activity Sensing is a crucial component of health monitoring and smart environment applications. Frequency Modulated Continuous Wave (FMCW) radars can be used for target tracking, but their collected data is usually accompanied by a significant amount of interference, especially in indoor environments hosting multiple human subjects, leading to a decrease in accuracy. In this paper, we propose a method that compensates that interference and can detect individual activities of multiple humans, overcoming existing methods’ limitation of detecting single human activities. To this end, a range-doppler map of the data is extracted with a FWCW radar, and the interference effect of this map is mitigated by a Convolutional Autoencoder (CAE). The CAE network learns to attenuate false positive regions to strengthen the target areas. This is followed by a Gaussian filter and then the targets are revealed by applying derivatives on both dimensions of the map. Evaluation results show that our method reaches activity recognition accuracies of 97.13% and 73.37% in the cases of one and two humans, respectively.