Abstract :

Human Activity Recognition plays a vital role in intelligent surveillance, healthcare monitoring, and human–computer interaction systems. This work proposes a robust HAR framework by combining ensemble learning of multiple Convolutional Neural Networks with MediaPipe-based pose estimation. MediaPipe extracts precise human skeletal landmarks that capture motion patterns effectively. These features are processed through multiple CNN models trained on diverse activity representations. Ensemble fusion enhances classification accuracy by reducing individual model bias. The system adapts well to variations in posture, lighting, and camera angles. Experimental evaluation demonstrates improved recognition performance compared to single-model approaches. The approach minimizes manual feature engineering through deep learning automation. Real-time processing capability makes the model suitable for practical applications. The framework supports scalable activity classes. Overall, the system del