UNSUPERVISED KEYPOINTS FOR REAL -TIME FALL DETECTION : COMPARATIVE ANALYSIS UNDER REAL - WORLD CONDITIONS WITH PREDICTIVE BANDWIDTH REDUCTION

Falls among older adults motivate reliable vision-based monitoring, but practical deployment is constrained by privacy, bandwidth, and partial body visibility. We investigate whether supervised anatomical pose is necessarily the best motion representation for fall detection under such real-world conditions. We propose a privacy-preserving framework using unsupervised motion keypoints, variational recurrent prediction, and sequence classification, and compare it against supervised OpenPose representations on the UR Fall Detection and Human Fall datasets under random, subject-disjoint, and occlusion-based evaluations. Under subject-disjoint evaluation with generally clear body visibility, supervised keypoints show a modest overall advantage. However, under controlled occlusion, unsupervised keypoints achieve 76.3% F1 compared with 56.3% for supervised pose, and retain 63.1% versus 38.4% F1 when half of the keypoint frames are forecast rather than transmitted. The occlusion advantage is statistically significant, with a median improvement of 21 F1 points. These results suggest that task-learned, anatomically independent representations can provide substantially greater robustness under deployment-relevant visibility degradation.