Human Activity Recognition (HAR) is an essential study area focused on identifying activities, especially in elderly care, where fall detection is vital for preventing serious injuries. mmWave radar sensors represent a non-contact and privacy-preserving alternative to traditional sensors. This study attempts to create a radar-based fall detection system that uses point cloud data and a deep learning approach based on PointNet. The primary focus is to address challenges such as limited dataset size, noise, and fluctuations in the number of points at each timestamp. These challenges are addressed through a pre-processing pipeline that includes augmentation, noise cleaning (ADBSCAN), data standardization (Moving Block Bootstrap), and temporal segmentation (sliding window). Classification is performed using the PointNet architecture, which directly processes point clouds. Based on testing results, the system achieved an accuracy of 99.97% with 0 False Positives and 2 False Negatives, as well as a Recall of 100% and 99.80% for the ‘not falling’ and ‘falling’ classes. This excellent performance, particularly the good recall for both classes and no False Positives, demonstrates the system's reliability. Supported by a comprehensive pre-processing pipeline, this success shows the system's great potential for improving the safety of the elderly through non-intrusive and accurate fall detection.