Human Action Recognition using Deep Neural Networks

In today’s industrial environments, safety supervision, workflow automation, and productivity rely heavily on accurate recognition of human actions. Existing methods often fail to capture the contextual dependencies between humans and objects, resulting in reduced performance in cluttered or dynamic settings. This paper proposes a Human–Object Contextual Interaction Network (HOCIN) that explicitly models spatial and temporal dependencies for improved action recognition. The framework integrates YOLOv8-based human–object detection with a context-guided attention mechanism, enabling extraction of discriminative cues from human pose dynamics and object affordances. Spatio-temporal features are fused with relational graph embeddings to form a unified representation for classification. The model is evaluated on two benchmark datasets—UCF101 and HAR— achieving recognition accuracies of 95% and 90% respectively, outperforming 3D CNN, RNN, LSTM, GNN, and two-stream baselines. These results demonstrate that HOCIN provides a robust and scalable solution for real-time monitoring in complex industrial environments. The originality of this work lies in its joint modeling of human motion and object context, bridging a key gap in existing action recognition approaches.