Advancing Abnormal Human Activity Recognition through Fused Heuristic Deep Learning and 3D Joint Angle Orientation Analysis

Recognizing abnormal human activity is critical in enhancing public safety, healthcare, and intelligent surveillance systems. This paper proposes a novel fused heuristic deep learning (FHDL) framework that integrates 3D joint angle orientation analysis with a heuristic reasoning module to improve the accuracy and interpretability of abnormal activity detection. By combining spatio-temporal features extracted through deep learning with domain-specific biomechanical rules, the proposed method addresses challenges related to occlusion, noise, and ambiguous motion patterns. Extensive experiments conducted on the NTU RGB+D 120 and Human3.6M datasets demonstrate that FHDL achieves state-of-the-art performance, attaining 99.4% accuracy on the Human3.6M dataset and an Area Under the Curve (AUC) of 0.98 on the NTU RGB+D 120 dataset. The framework significantly outperforms traditional Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Spatio-Temporal Graph Convolutional Network (STGCN), and transformer-based models. Ablation studies and statistical significance tests validate the complementary contributions of the heuristic module and joint angle features. The results suggest that integrating handcrafted knowledge with data-driven models significantly enhances abnormal activity recognition performance in complex environments.