This extensive review aims to provide an in-depth analysis of the latest free-living human activity recognition (FL-HAR) methods and their applications in health research. As sedentary lifestyles become more and more common, accurate and scalable technology for detecting human activity in free-living situations is crucial for enhancing public health programs. The aim of this research project is to collect high-quality accelerometer data from a range of strategically placed sensors on the human body. After improving the raw accelerometer signals through pre-processing, relevant data representing different activities is extracted using feature extraction algorithms. A range of state-of-the-art machine learning algorithms, including deep learning models, are used to train and optimize the recognition system. Furthermore, the research delves into potential applications of the created HAR system in many domains such as human-computer interaction, sports surveillance, and medical area. In view of this, the review paper focuses on new models for machine learning in the temporal domain and hybrid feature selection methods. The article examines several benchmark datasets used in current methods. The report concludes with a list of problems and concerns that need more investigation and development.