Human Activity Recognition is now recognized as a significant area of research, with applications in medical, smart environments, and human-computer interaction. This paper reviews the latest developments in these systems, focusing on sensor technologies, edge computing, and ML frameworks. The increasing use of wearable devices, such as smart mobile phones and smartwatches, has facilitated the collection of diverse datasets, including acceleration and motion data that is crucial for activity classification and fall detection. Traditional approaches to HAR typically rely on feature engineering and labeled datasets. Experimental results show a 30% reduction in latency and a 25% decrease in power consumption while achieving a classification accuracy of 95.7%. This paper examines various sensing modalities, from motion sensing units to vision-based systems, as well as emerging modalities like millimeter-wave radar. Lightweight neural networks and on-device carry-forward learning are examples of contemporary techniques that allow real-time HAR on resource-constrained tech and model personalization. Emphasis is given to the differences in the interpretation of the same activities, such as varying categorizations for falling. Processes like semi-supervised learning, multimodal fusion, and dynamic time warping enhance recognition performance and adaptability. Although many HAR systems achieve high recognition rates of 90–99% across various tasks, significant challenges remain. These include the concept of drift, low power consumption, limited availability of labeled data, and individual differences in usage patterns. However, adaptive architectures, unsupervised learning, and multimodal integration are promising strategies that form the way for more robust and scalable solutions.