Optimizing Human Activity Recognition Using Stacked CNN with Ensemble Learning

In recent times, Human Activity Recognition (HAR) plays a crucial role in numerous artificial intelligence applications, including surveillance, healthcare, and security. However, the primary challenges in accurately identifying human activities from video data include occlusion, low resolution, data scarcity, and high computational costs. The main objective of this study is to enhance both the recognition accuracy and computational efficiency. This study presents a computationally efficient HAR model that leverages deep learning and machine learning algorithms for feature extraction and classification tasks. The proposed model consists of three fine-tuned convolution neural networks: VGG16, VGG19, and MobileNetV2 for feature extraction, combined with four classifiers, Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), and Fuzzy classifier. The performance of the proposed model was evaluated using three benchmark datasets: KTH, YouTube11, and Peliculas. The proposed method demonstrated exceptional performance of 99.60% and 99.50% accuracy on the KTH and Peliculas dataset using MobileNet and SVM. Additionally, the proposed method achieved a superior performance of 99.80% on the YouTube11 dataset with a combination of VGG16 and SVM. Finally, the performance of the proposed model was compared and validated against previous models.