Egocentric 3D Skeleton Learning in Identity-Aware Deep LSTM Network Encodes Obese-Like Motion Representations

Recent advancements in 3D motion capture technology are emerging as a crucial catalyst for future developments in healthcare. With obesity increasingly recognized as a significant health concern stemming from poor dietary habits, our research focuses on identifying early indicators of obesity-inducing dietary patterns using 3D time-series skeleton data. Initially, we gathered 3D time-series skeletons from mouse models with diet-induced obesity. Subsequently, we explored the effectiveness of different viewpoints for analyzing 3D skeleton data: allocentric versus egocentric perspectives. Finally, we sought to develop efficient deep recurrent network architectures. Our findings demonstrate that integrating the concept of an egocentric viewpoint into 3D skeleton data analysis, coupled with training deep LSTM networks to accurately classify identities, can effectively distinguish motion differences induced by diet between control and high-fat diet groups. This research offers a viable approach to leveraging deep learning for early detection of health risks, facilitating timely interventions and broadening the scope of healthcare technology.