Human Activity Recognition with EfficientNetB3: A Model-Driven Approach to Complex Activity Classification

Human Activity Recognition (HAR) is an application that encompasses surveillance to health care monitoring. The paper propounds deep learning (DL) in the classification of 15 different activities including calling, cycling, dancing, and running etc. through the use of EfficientNetB3 for the convolutional neural network. The dataset comprises more than 12,000 images with one specific category per image. In the experiment, different forms of data augmentation were used to avoid overfitting and enhance generalization. EfficientNetB3 has excellent balance between accuracy and computational efficiency in handling intricate complex patterns with fewer parameters and thus adopted for this experiment. The model was trained from scratch for 50 epochs, resulting in an accuracy of 90.10%, which was higher than that of traditional and hybrid methods of HAR. From a comparative analysis, the proposed model outperforms all competing approaches, such as YOLOv5 with LDA and SVM, as well as hierarchical classification models. The results show that DL based models work very well in discrimination of complex human activities. In an attempt to further make the system more versatile, future work will concentrate on optimization of the model for real-time HAR, fusion with sensor data, and extension to larger datasets with more classes of activities.