Residual DCRNN-based physical contact estimation from human motion

초록

Understanding physical contact between the human body and surrounding objects is a critical factor in scene interpretation and human-object interaction (HOI) analysis. While prior approaches primarily rely on visual or depth data, they remain highly sensitive to environmental conditions such as occlusion and lighting, limiting their generalizability. To overcome these challenges, we propose Residual Diffusion Convolutional Recurrent Neural Network (R-DCRNN), a deep learning architecture that estimates frame-wise contact regions solely from 3D skeletal motion sequences. The model integrates three core modules: Residual Graph Convolution to capture anatomical joint structure, Joint Attention Pooling to emphasize contact-relevant joints, and a GRU (Gated Recurrent Unit)-based temporal encoder to model sequential dynamics. To improve temporal consistency, a Temporal Smoothing technique is applied in post-processing. Extensive experiments show that R-DCRNN outperforms both non-graph models (e.g., Long Short Term Memory(LSTM)) and graph-based models (e.g., Spatio-Temporal Graph Convolutional Network(ST-GCN)) across multiple evaluation metrics, including accuracy, precision, recall, and F1-score. Furthermore, it achieves the lowest frame-to-frame variation and smoothness loss, indicating robust temporal stability. Visualization results using a 3D character simulation confirm the model’s ability to localize contact regions reliably, underscoring its potential for real-time applications in Virtual Reality/Augmented Reality(VR/AR), human-robot interaction, and behavioral analysis. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2026.

키워드