Human activity recognition (HAR) through vision-based systems has gained significant attention due to its wide-ranging applications in various domains such as healthcare, security, and human-computerinteraction. Convolutional Neural Networks (CNN) have emerged as powerful tools for feature extraction and classification in visual data processing tasks. Additionally transfer learning models on large datasets, has shown promise in enhancing the performance of HAR systems, particularly when labeled data is limited. This paper presents a comprehensive review and analysis of leveraging CNN architecture and transfer learning methods for vision-based HAR. We investigate the application of transfer learning in HAR, wherein features learned from large-scale datasets are transferred to HAR tasks with smaller, domain-specific datasets. This paper aims to provide insights into the utilization of CNNs and transfer learning techniques for advancing the field of vision-based human activity recognition, ultimately contributing to the development of more efficient and reliable HAR systems