Human activity recognition: A review of deep learning approaches
Somanshu Gupta, Abhilasha Sharma · 2025
The ability to interpret human body gestures or motions as well as determine human activity or other actions is termed Human Activity Recognition (HAR). Increasing utilization of HAR across several applications including healthcare, surveillance, sports, and elderly care has led to tremendous advances in research activity. This article systematically reviews various deep learning (DL) approaches towards enhancing HAR that includes Convolutional Neural Networks (CNNs), Deep Convolutional Neural Networks (DCNNs), Deep Neural Networks (DNNs), and Recurrent Neural Networks (RNNs). The method applied entailed extensive reading of scientific literature for a comparative evaluation of the performance, limitations, and applicability of the mentioned DL models to solve HAR problems including human pose estimation and complex spatiotemporal pattern discovery. The observations point out that CNN-based techniques excel in identifying local temporal relations, while RNNs excel at sequential data modeling but at the cost of vulnerability to the problems of capturing long-term dependencies. Despite impressive advances in performance realized by such approaches, challenges such as inadequate labeled data, scalability issues, and issues experienced in outdoor environments continue to exist. The review not only outlines the current state of advancements in HAR but also finds out the loopholes in research and suggests potential directions of improvement towards making HAR systems robust and scalable.