A Survey on Human Action Recognition: Deep Learning and Machine Learning

Human action recognition systems are essential technologies used in many applications, including monitoring older adults and patient activities in healthcare, improving security systems by monitoring suspicious behavior, analyzing sports performance, and supporting human-machine interaction. This paper examines the evolution of human action recognition methods from traditional machine learning approaches to modern deep learning techniques in recent years. DL has shown significant progress in dealing with complex datasets and overcoming machine learning limitations by replacing handcrafted feature extraction with modern methods for automatic feature extraction. DL models such as 3D convolutional neural networks and bi-directional long short-term memory have shown remarkable progress in performance, achieving accuracy higher than 97.98% on benchmark datasets such as HMDB-51 and UCF101. This success has been demonstrated in different types of datasets, such as video and skeleton data, reflecting the flexibility and robustness of deep learning methods. However, high computational requirements and incomplete or noisy data remain open issues. To solve these challenges, future research may be moving towards employing ensemble and fusion techniques that integrate more than one type of data, such as video, skeleton, and depth data, with the addition of advanced techniques, such as attention mechanisms, to improve feature selection and emphasize relevant information. These approaches and techniques can improve the model's accuracy, robustness, and efficiency. They provide a better and clearer understanding of the performing of movements to distinguish human actions, overcome their limitations, and support the development of real-time applications.