A Systematic Literature Review of Optimization Methods in Skeleton-Based Human Action Recognition

Skeleton-based human action recognition (HAR) has become a prevalent topic in the field of computer vision due to its ability to widely use in different applications such as video surveillance, healthcare monitoring, sport analysis, and human computer interaction. Over the past decade, researchers have introduced various approaches such as CNN, RNN, and GCN to classify human actions using skeleton data. However, challenges such as noisy data, inter-class similarity, and diverse action dynamics require more robust and efficient solutions for skeleton-based HAR. To address these challenges, optimization methods, such as attention mechanisms and fusion strategies are being implemented in the feature extraction and classification processes to enhance the performance of action recognition. Despite their critical role in improving the recognition accuracy and system reliability, there is currently no systematic review that focuses specifically on these optimization methods in the context of skeleton-based HAR. This study presents a systematic review of the current optimization methods by thoroughly examining 92 publications published from 2014 to 2024. This study aims to provide the researchers with a better understanding of how optimization methods are integrated in recent HAR systems and serve as a reference for future studies on improving skeleton-based HAR.