Multi-stream activated adaptive graph network based on incomplete skeletons for thermal adaptive behavior recognition

Research on video-based thermal comfort prediction has grown rapidly, yet existing methods suffer from severe performance degradation under skeleton occlusion and incomplete joint detection in real indoor surveillance scenarios. To address this issue, this paper proposes a multi-stream activated adaptive graph network (MTAGCN) for thermal adaptive behavior recognition based on incomplete skeleton sequences. The proposed model constructs a Transformer-enhanced adaptive graph convolution module to dynamically capture spatial-temporal topological correlations among human joints. Three synthetic occlusion datasets including frame occlusion, partial body occlusion, and random joint occlusion are established to evaluate model robustness. Experimental results show that MTAGCN outperforms baseline methods on both the original thermal adaptation behavior dataset and occlusion-contaminated benchmarks, achieving superior recognition accuracy and stronger robustness against skeleton incompleteness. This work provides a low-cost, non-intrusive visual sensing framework for dynamic thermal comfort perception, and lays a foundation for intelligent indoor environmental regulation integrating behavioral recognition and thermal comfort prediction.