HNLS: Designing a Robust Model for Human Action Recognition Based on Postures by Using Hybrid Neural Learning Strategy

Human action recognition is a pivotal area in computer vision with numerous applications, such as video surveillance, healthcare monitoring, sports analysis, and human-computer interaction. This paper presents a robust methodology, Hybrid Neural Learning Strategy, which integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks for effective human action recognition. The proposed approach utilizes the UCF101 dataset, which captures 101 diverse human actions with posture-specific information under various lighting, angles, and environmental conditions. The data preprocessing steps include cleaning, feature extraction using OpenPose, and dimensionality reduction through AutoEncoders. A hybrid architecture combining CNN for spatial feature extraction and LSTM for temporal dependency modeling is employed. The model achieves an accuracy of 93.78%, outperforming traditional and hybrid models, such as CNN, RNN, and AutoEncoder-LSTM models. These results demonstrate the model's efficacy in recognizing complex human actions across various scenarios, making it highly suitable for applications in surveillance, healthcare, and human-computer interaction.