A Deep Learning Approach for Human Activity Recognition in Healthcare Applications

Abstract

Human Activity Recognition (HAR) has become an important area of research
due to its wide applications in surveillance, healthcare, and human computer interaction, where
accurate understanding of human behavior is essential. However, traditional approaches and
basic convolutional neural networks often face limitations in capturing complex visual patterns
and require large amounts of data for effective performance. This project aims to develop an
efficient and accurate HAR system by leveraging transfer learning with the EfficientNetB0
model and comparing its performance with a conventional CNN approach. The proposed
methodology utilizes EfficientNetB0 as a pre-trained feature extractor to learn high-level spatial
features from input images or video frames, followed by fully connected layers for classification,
while the existing CNN model is trained from scratch for performance comparison. Experimental
results demonstrate that the proposed EfficientNetB0-based model achieves higher accuracy,
compared to the CNN model, while also reducing training time and improving generalization.
The study concludes that integrating transfer learning with EfficientNetB0 significantly enhances
the effectiveness of human activity recognition systems, making it a reliable and scalable
solution for real-world applications.