Graph Attention and Contrastive Learning for Enhanced Spatio-temporal Human Activity Recognition

Human activity recognition (HAR) is a core component of intelligent surveillance, assistive healthcare, and human computer interaction.Fine-grained recognition, however, remains challenging due to subtle inter-class similarities, viewpoint changes, and variability in motion dynamics.Graph convolutional network (GCN) based methods model local joint relations effectively but often under-represent long-range temporal structure, while purely transformer based models improve sequence modeling at the cost of heavier computation.This work presents a Graph-Attentive Contrastive Transformer for Human Activity Recognition (GACT-HAR) that combines graph-attentive spatial encoding, transformer based temporal aggregation, and an augmentation driven contrastive objective to refine feature discriminability.The key contribution lies in the joint use of adaptive graph attention and contrastive representation learning on skeleton sequences, which strengthens separation between closely related actions while preserving temporal coherence.Evaluation on the NTU-RGB+D-120 benchmark, comprising 120 action classes, 106 subjects, and roughly 114,480 three dimensional skeleton sequences, shows that GACT-HAR attains 95.3% Top-1 accuracy, 94.9% macro F1, and 172 frames per second with about 5.1GB GPU memory.This corresponds to an absolute Top-1 gain of up to 7.3 percentage points over recent baselines, indicating suitability for real time HAR in healthcare, robotics, and large scale video analytics.