An Efficient Human Activity Recognition Scheme using Hybrid Deep Learning Strategy with Intelligent Classification Principles

Human Activity Recognition (HAR) is a critical component in various domains, including healthcare, security, and human-computer interaction. Accurate and efficient HAR systems are essential for understanding human behaviors and facilitating smart applications. In this study, we present a novel approach for HAR using a hybrid deep learning strategy that combines Convolutional Neural Networks (CNN) and Light Gradient Boosting Machines (LightGBM) with intelligent classification principles. We evaluate the proposed scheme on the publicly available Stanford40 Human Action Dataset, demonstrating its effectiveness and versatility. The Stanford40 Human Action Dataset contains a diverse range of human activities captured through sensors and cameras. Our hybrid approach begins by preprocessing and feature extraction from this dataset, aiming to extract informative features that capture both temporal and spatial patterns. Our hybrid strategy involves a two-stage classification pipeline. In the first stage, a CNN-based model is employed to learn intricate representations from the feature vectors. In the second stage, LightGBM, a powerful gradient boosting framework, is introduced to refine the classification process. LightGBM focuses on addressing misclassified instances from the CNN's output, acting as an intelligent ensemble component. This combination enhances overall recognition performance significantly. Experimental results demonstrate the superiority of our hybrid approach with an accuracy of 96%, when compared to traditional methods and single-model architectures. This approach combines the strengths of deep learning and gradient boosting while introducing adaptability and context-awareness to achieve accurate and reliable recognition of human activities, thereby contributing to the development of smarter and more responsive human-centric technologies.