Real-time human activity recognition using deep learning techniques for the next-generation healthcare system

Abstract

In today's world, healthcare systems are being built with human activity recognition (HAR) to help the elderly, disabled, and children's activities by constantly observing their behavior. However, HAR using computer vision and traditional machine learning techniques is not an efficient use of healthcare system resources because of potential and accuracy issues. This study aims to examine the use of deep learning techniques in real-time HAR. This proposal is a hybrid method that utilizes the EfficientNetB0 architecture and a convolutional extension of a long short-term memory network (EfficientNetB0ConvLSTM), to achieve human-like intelligence. The EfficientNetB0 is utilized to extract image features, and convolutional long short-term memory (ConvLSTM) is utilized to categorize six human actions to recognize human activities. This approach leverages the strengths of convolutional neural networks (CNNs) in extracting spatial features from video frames and LSTMs in capturing temporal dependencies within activity sequences. Firstly, an extensive investigation is conducted on existing literature studies to select a suitable dataset. Next, the proposed method was evaluated on the challenging HMDB51 video datasets and finally achieved an accuracy of 89.22%, which is significantly higher than other methods on this dataset. This outcome shows the potential of EfficientNetB0ConvLSTM for real-time HAR applications like healthcare.