Modeling transformer architecture with attention layer for human activity recognition

Human activity recognition (HAR) is necessary in numerous fields, involving medicine, sports, and security. Traditional HAR methods often rely on complex feature extraction from raw input data, while convolutional neural networks (CNN) are primarily designed for 2D data. The proposed approach seeks to overcome these limitations by leveraging both spatial and temporal attributes for improved action detection and enhancing the understanding of human movements across adjacent frames. This research aims to address the challenges of HAR by introducing a new model that combines a 3D CNN architecture with an attention layer. A 3D convolution transformer is employed to capture intricate spatial and temporal features, generate multiple data channels from input frames, and optimize performance through regularization and model ensemble techniques. The main findings reveal outstanding results on benchmark datasets, with an accuracy of 98.09% and 99.09% on the Weizmann and UCF101 datasets, respectively. These results underscore the model's effectiveness in accurately identifying human activities in movie-based natural environments.