The performance of the skeleton-based human activity recognition (HAR) system heavily relies on robust datasets. Due to the time-intensive process of manual segmentation and annotation, researchers often face challenges in creating a flawless dataset. Consequently, there are a limited number of publicly accessible skeleton-based HAR datasets, which are limited in scope, impacting research progress. In order to overcome these limitations this work suggests an automatic segmentation technique that effectively determines activity start and end points leveraging reconstruction error between frames. This method significantly simplifies the dataset preparation process eliminating the need for manual segmentation. In contrast to traditional PCA-based methods that fail to recreate the original data, the proposed method integrates an autoencoder-based framework, which preserves original feature embeddings. Following segmentation, a novel HAR system was developed that enhances overall HAR efficiency by combining activity segmentation and recognition. In order to identify segmented but unlabelled actions and facilitate subsequent analysis without the need for human intervention, a deep learning-based clustering technique is presented. The proposed method for activity segmentation was evaluated on three publicly accessible unsegmented HAR datasets, while the recognition method was tested on five datasets. The proposed segmentation method achieved an accuracy of over 88% on all datasets, which is a significant improvement over the existing method. On a large-scale HAR dataset, activity recognition accuracy is improved by 2%, and the clustering method demonstrated a remarkable 24% performance improvement compared to existing traditional clustering techniques while performance gain over existing deep learning-based clustering methods was over 11%. These advancements demonstrate the robustness of the proposed approach. As part of this research, two real-world systems were developed which further contributed to practical applications. Firstly, a cost-effective social distance monitoring system that used 3D skeleton data to estimate the distance between individuals in real time was developed. The system achieved over 98% detection accuracy, effectively identifying social distancing violations, and providing a scalable solution to assist enforcement agencies. The second method is an athlete performance evaluation system utilizing 3D skeleton joint tracking to analyse dynamic movements such as jumping jacks. Through the quantification of joint movement over 500 frames, the method offered comprehensive insights into performancevariation and consistency between athletes, which has significant potential applications in sports science, biomechanics, and rehabilitation. This research addresses critical obstacles in dataset preparation and activity recognition by introducing innovative approaches to the HAR domain. A comprehensive solution that enhances the efficiency and accuracy of HAR systems is presented in this work, through the integration of automatic segmentation, clustering and recognition, into a unified framework, while demonstrating impactful applications in public health and sports performance analysis. The findings of the work have practical implications for large-scale HAR applications, particularly in areas such as healthcare, sports analysis, and surveillance, where automated and accurate activity recognition is essential.