Abstract
With the recent advances in multi-disciplinary human activity recognition techniques, it has become inevitable to find an efficient, economical & privacy-friendly approach for human-to-human mutual interaction recognition in order to breakthrough the modern artificial intelligence centric indoor monitoring & surveillance system. This study initially attempted to set its sights on the already proposed human activity recognition mechanisms and found a void in mutual interaction recognition from Wi-Fi channel information which is convenient & affordable to be utilized. Then it elucidated on the corresponding components of wireless local area network gadgets along with the channel properties, and notable underlying causes of signal & channel perturbation. Thenceforth the study conducted three experiments on human-to-human mutual interaction recognition using the proposed Self-Attention furnished Bidirectional Gated Recurrent Neural Network deep learning model which is perceived to become emergent nowadays for time-series data classification through automated temporal feature extraction. Single pair mutual interaction recognition experiment achieved a maximum of 94% test benchmark while the experiment involving ten subject-pairs secured 88% benchmark with improved classification around interaction-transition region. Demonstration of a graphical user interface executable software designed using PyQt5 python module subsequently portrayed the overall mutual human-interaction recognition procedure, and finally the study concluded with a brief discourse regarding the possible solutions to the handicaps that resulted in curtailments observed in the case of cross-test experiment.
Abstract (translated)
URL
https://arxiv.org/abs/2202.08146
