Contrastive Representation of Channel State Information for Human Body Orientation Recognition in Interaction with Machines

Human activity recognition using WiFi signals is a recent trend for interaction with smart devices in indoor environments. In a smart environment with multiple smart devices, orientation recognition of the user while performing an activity enables identification of the device which the user intends to interact with. Wireless signals are uninterpretable by human and the large-scale collection and annotation of wireless signals for supervised learning is very challenging. In this paper, we propose contrastive learning (CL) from WiFi signals for human orientation recognition (HOR) while performing an activity. Results show that a hierarchical unsupervised CL model trained on a large-scale unannotated dataset of WiFi signals and then jointly fine-tuned with a classifier on an external small-scale annotated dataset has a high classification performance. This approach is particularly important in cross-domain adaptation of HOR systems.