Unsupervised keypoints beat supervised for fall detection under occlusion
Supervised pose estimation misses half of falls; unsupervised keypoints stay robust.
A new study from Tasmiah Haque, Jacob Kosinski, Sumit Mohan, Srinjoy Das, and Mohammad Abdullah Al-Mamun tackles a critical safety challenge: detecting falls in older adults without compromising privacy or bandwidth. Traditional supervised pose estimation requires transmitting full RGB video and relies on visible anatomical landmarks—making it fragile when bodies are partially occluded by furniture, blankets, or the fall itself. The team proposes a privacy-preserving framework that processes video locally, extracting unsupervised keypoints and using variational recurrent prediction to detect falls from both observed and forecasted motion patterns.
Evaluating on the UR Fall Detection and Human Fall datasets under random, subject-disjoint, and occlusion-based splits, the results were striking. Under occlusion, supervised keypoints missed nearly half of all falls, while unsupervised keypoints retained strong sensitivity and produced fewer false positives for complex activities. The gap widened under bandwidth constraints, where supervised localization errors compounded through the temporal model. Unsupervised keypoints adapt to visible body structure rather than fail on absent landmarks, making them a better choice for real-world deployment where partial visibility is common.
- Unsupervised keypoint framework operates locally, transmitting only compact motion representations instead of RGB video to preserve privacy and reduce bandwidth.
- Under occlusion-based evaluation, supervised keypoints missed nearly half of all falls, while unsupervised keypoints maintained strong sensitivity.
- Subject-disjoint evaluation showed supervised keypoints performed better only when all anatomical landmarks were fully visible.
Why It Matters
Practical fall detection systems for elderly care can now work reliably under real-world occlusion, preserving privacy and bandwidth.