Privacy invasion
Concern during intimate-activity monitoring
See the motion.
Keep the privacy.
Enabling Pose-Only IoT Cameras
for Privacy-Preserving Edge Vision
The University of Hong Kong
Privacy is the barrier.
Video-based care can capture more than an activity.
Data misuse
Concern during intimate-activity monitoring
Mean agreement, 1–6 scale (higher = stronger concern). Online survey, N=146. Maidhof et al., 2023 ↗
Identity
Faces · Expressions
Intimacy
Clothes · Personal care
Home
Family · Private spaces
Keep RGB local.
Share only the pose.
Less appearance exposure. Motion can still be sensitive.
Only the skeleton leaves.
On-device vision. Compact output. Useful activity sensing.
107 ms
p95 pose latency
STM32H753 MCU76 bytes
One skeleton record
12 joints · fixed-format SFSK248.64 KiB
Peak activation memory
Model activations onlyHow it works.
📷 Local RGB → 🦴 Skeleton stream → 🚶 Activity recognition
On the camera Trusted boundary
Capture
Local RGB frames
→
Estimate & refine
Compact pose model
Raw frames stay local and ephemeral
76 B→SFSK record
12-joint skeleton
Geometry, without RGB
↓
Downstream activity recognition
StandingSittingMoving
Memory-aware
Narrow early stages · Expand after downsampling
Motion-aware
Visual features + Arm geometry → Refined joints
Scene-aware
One empty-scene snapshot · Wake on motion
Results.
Two benchmarks. One real-world deployment.
NTU RGB+D
FP32 evaluation78.40%
Pose PCK@0.10
74.99%
HAR Top-1 accuracy
±1.06% · three-seed meanPenn Action
FP32 evaluation77.18%
Pose PCK@0.10
87.46%
HAR Top-1 accuracy
±0.12% · three-seed meanFP32 pose · HAR mean ± s.d., three seeds.
In a real room.
~40 minutes · One participant · Four activity states
96.45%
Activity macro-F1
3.15%
Mean absolute activity-duration error
Measured on the annotated session.
Three days.
Everyday motion.
🏠 Fixed camera · One older adult
☀️ Three daytime monitoring sessions
state transitions / hour
state transitions / hour
state transitions / hour
From pose streams to daily activity patterns.
Explore the case study ↗Descriptive monitoring results; no dense ground-truth labels. Accuracy is evaluated in the separate annotated session above.
Built for
the camera.
STM32H753 + OV2640 · On-device pose extraction
Full inference
when motion starts.
💤 Inactive → Low-power checks
🚶 Motion → Pose inference
~63%
Lower MCU-side current
during inactive periods
STM32H753
- Processor
- Arm Cortex-M7 · 480 MHz
- Image sensor
- OV2640
- Pose model
- 294.3K parameters
- Computation
- 29.56M MACs / frame
StickFigure
Enabling Pose-Only IoT Cameras
for Privacy-Preserving Edge Vision