This paper proposes a novel approach for detecting falls in elderly individuals using a vision-based system enhanced by contrastive learning techniques. Traditional fall detection models often struggle with high false positive rates and generalization issues in diverse environments. By leveraging contrastive learning, we aim to improve the feature representation of fall-related events and non-fall actions, ensuring better discrimination between the two. Our model utilizes 3D vision-based body articulation data, feeding it into a contrastive learning framework that learns robust feature embeddings for various human activities. The model learns a low-dimensional embedding of the head position based on spatial coordinates, enabling robust estimation and generalization even with incomplete or missing data during testing. The experimental results on a standard fall detection dataset demonstrate that our approach achieves state-of-the-art performance, reducing false positives while maintaining manageable accuracy in real-world settings.