SKELETON-BASED HUMAN ACTION RECOGNITION USING TRANSFORMER MODEL WITH SOFTMAX WITH MULTIDIMENSIONAL CONNECTED WEIGHTS

Abstrak

Skeleton-based human action recognition (HAR), particularly from CCTV surveillance footage, has garnered significant interest within the artificial intelligence community. The skeletal modality provides a robust, high-level representation of human motion. Prevailing methods in this domain predominantly rely on a joint-centric approach, modeling the human body as a set of coordinate points. However, this representation often fails to fully capture the rich structural and kinematic relationships essential for accurate motion classification. To address this limitation, we propose a novel method termed SoftMax with Multi-Dimensional Connected Weights. This approach enhances classification by explicitly modeling the informative connections between body joints, represented as skeletal edges. We develop an end-to-end deep learning framework that learns discriminative spatio-temporal representations directly from sequences of skeleton point vectors using Convolutional Neural Networks (CNNs). Results demonstrate that our approach achieves stateof-the-art performance, underscoring the effectiveness of leveraging skeletal edge information and advanced classification techniques for human action recognition.