JLHI dataset: a joint localization and human-to-human interaction dataset for indoor environments with different...

This paper introduces a Wi-Fi based joint localization and human-to-human interaction recognition dataset across diverse indoor layouts, known as the JLHI dataset. It includes 5 distinct interactions performed by 20 different pairs of subjects at 3 locations in 3 different indoor layouts with varying occlusion levels, ranging from open spaces to highly cluttered environments. Each pair of subjects performed 50 trials for each interaction per position per layout, resulting in a total of 45,000 trials (i.e., 20 pairs of subjects x 5 interactions x 50 trials x 3 locations x 3 layouts). To generate and collect the dataset, NI USRP-2901 software-defined radios operating at 5.2 GHz with a sampling rate of 5 MHz were used. The capture Wi-Fi signals contain channel state information with including amplitude and phase across 64 subcarriers, along with received signal strength indicator values, recorded using a custom CSI logger coded using GNU Radio Companion. Unlike existing Wi-Fi based human activity datasets which predominantly focus on single-person activities or human-to-human activities in a fixed layout or position, our dataset bridges this critical gap by providing a multi-user interaction dataset under diverse spatial positions and occlusion conditions. It enables the design of robust machine learning-based models for joint localization and human-to-human activity recognition in dynamic indoor environments that are resilient to environmental changes. Additionally, the dataset offers a realistic benchmark for evaluating model effectiveness across diverse layouts, spatial positions, and occlusion levels, enabling the development of systems that perform reliably in real-world settings.