Model
Pose Estimator
Extract body keypoints to support intent, posture cues and gait sampling.
beta
Body Keypoints
KeypointsGait
Overview
A keypoint model runs periodically (every few frames) on tracked people. The skeletons feed behavioural intent estimation, posture / fall cues, and the gait descriptor that contributes to cross-camera re-identification. Running it on an interval keeps it off the per-frame critical path.
| Role | Body Keypoints |
|---|---|
| Expected latency | ~5–12 ms / person (GPU) |
| Hardware requirements | NVIDIA GPU recommended |
| Memory requirements | ~1–2 GB VRAM |
Inputs & Outputs
Inputs
- Person crop / frame
- Track id
Outputs
- 2D keypoints per person
- Gait descriptor contribution
Advantages & Limitations
Advantages
- Enables intent + gait
- Interval scheduling caps cost
- Robust to partial visibility
Limitations
- Crowded overlap confuses keypoints
- Needs enough resolution on the subject
- Beta accuracy on unusual poses
Example outputs
Pose Estimator
A walking skeleton feeding a gait signature
Pose Estimator
Posture cues for intent estimation
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
pose.enabled | true |
pose.interval | 5 |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| Keypoint AP | COCO (ref) | AP | ~68 | reference |