Radar-Based People Counting Under Heterogeneous Clutter Environments

Radar-based people counting (RPC) systems, which perceive their surroundings through wireless reflections from radar, can provide crowd information free from privacy invasion and illumination problems. Current RPC approaches leverage a deep supervised learning framework to classify the number of individuals from complicated radar data, based on a labeled set of signals. However, they have a critical limitation in terms of practicality in that even a marginal change in the surrounding environment requires new data acquisition and time/labor-consuming crowd labeling. To address this problem, this study proposes a novel framework that transfers the RPC semantics from a previous dataset constructed under a certain environment into reflections from other environments, thus enabling fully unsupervised RPC in variable clutter environments. Considering that the main sources of environmental bias originate from environment-variant components within the radar signal itself and the embedded feature domain, our key idea is to mitigate the environmental dependency in each aspect through radar-centric stochastic signal augmentations and unsupervised domain adaptation. This process allows the model to confuse environment-dependent signatures and focus solely on the reflections from humans. Based on experiments under real measured datasets, we verify the robustness of the framework in dynamic clutter conditions, where existing RPC approaches completely fail.