AI Edge Rehabilitation Posture Analysis

Edge cameras track rehabilitation posture and detect compensation on site.

Real-time clinical biomechanics kinematics, 33-point BlazePose INT8 tracking, phase-state machine, and debounced compensation detection on reCamera edge devices with zero video leakage.

SenseCraftBeginner5minVision & Video AIAI Agent ReadySource Code

reCamera Pro live running view — 33-keypoint full-body skeleton, real-time ROM angle and repetition counter

Real-time squat rehab evaluation — knee flexion ROM and valgus compensation alert

Edge system architecture — smart camera outputs WebSocket primitives and RTSP video with zero leakage

reCamera Pro hardware and clinical camera placement (2.0–3.5m front view)

reCamera 2002 cost-efficient hardware preset (RISC-V 1GHz TPU camera)

Reusable Site Templates

Indoor fixed views, patients facing the camera, needing guided movements and instant correction.

Physical therapy and rehab clinics

Outpatient rehab rooms and kinesiology gyms, assisting therapists monitoring multiple patients.

Senior living and daycare centres

Senior activity halls and sit-to-stand training corners with high-contrast HUD cards.

Home rehabilitation

Living rooms and home training areas. Power up the all-in-one camera with complete patient privacy.

Not a fit: steep overhead views, distance beyond 4 m, heavy occlusion.

How it works

The all-in-one smart camera runs pose estimation and kinematics on its built-in NPU/TPU, emitting vector primitives over WebSocket and RTSP while video stays on device.

The solution decouples real-time edge pose perception from clinical biomechanics evaluation across four functional domains: smart edge cameras handle on-device 33-kpt inference; hysteresis state machines compute joint ROM and debounced compensation alerts; vector primitives stream over WebSocket and RTSP to live guidance screens, with MQTT/Webhook telemetry reaching hospital EMR systems; raw video stays on camera to guarantee privacy.

reCamera Pro and reCamera 2002 carry the image sensor, lens, and edge accelerator in one single unit.

Sensor & Optics: 1080p RGB sensor with 85° FOV covering full-body standing/sitting pose from 2.0 to 3.5 meters.
Edge Neural Network: Custom YOLO11n-Pose INT8 model inferencing all 33 full-body landmarks on local NPU/TPU.
Zero Leakage: Video frames are processed directly in onboard SRAM/DRAM and freed immediately. No video stream crosses the network interface.

Clinical kinematics state machine tracks Range of Motion (ROM) and pathological compensations:

ROM Angles: Continuous Euclidean 3D joint angle tracking (knee flexion 0°–180°, shoulder abduction 0°–180°, elbow flexion).
Phase Transitions: Hysteresis state machine (READY → ASCENT → INFLECTION → DESCENT → COMPLETED) eliminates partial-rep false counts.
Compensation Detection: 3-frame temporal debounce verifies knee valgus (>8°), shoulder shrug (>10°), or excessive trunk lean.

Clean decoupling between on-camera evaluation and display hosts:

WebSocket 8080: Emits Result Hub v2 geometry[] vector primitives and real-time HUD metrics JSON.
RTSP 8554: Optional H.264 stream with hardware OSD burned-in for standard hospital NVR archiving.
HTTP 8080: Local Web App Center console for live parameter tuning without recompilation.

Engineering details, test methodology and measured dataRead the engineering wiki →

How well it works

What patients and therapists see on the local display — the live rehabilitation HUD.

33-Keypoint skeleton and HUD dashboard

Real-time 33-joint skeleton topology overlay
Large cards showing phase and valid repetition count
Dynamic joint ROM angle tracking with target guidelines

Clinical kinematics and compensation alert

Continuous joint Range of Motion measurement
Red highlighted alert on knee valgus or shoulder shrug
3-frame debounce prevents partial-rep counting

Dual hardware presets and zero video leakage

reCamera Pro achieves 24.0 FPS INT8 inference
Video discarded from RAM, never leaving camera
Streams WebSocket JSON primitives and RTSP video

Measured results

Engineering benchmark on clinical rehabilitation videos — assistive movement guide, not medical diagnosis.

Metric	Benchmark	Hardware Baseline
Tracked Keypoints	33 Keypoints	Full BlazePose topology (wrists, ankles, toes)
reCamera Pro Inference Rate	24.0 FPS (41.7 ms)	RV1126B NPU, INT8 quantized
reCamera 2002 Inference Rate	13.3 FPS (71.1 ms)	SG2002 TPU, hardware-fused INT8
ROM Angle Measurement Error	< 2.5°	Calibrated against clinical goniometers
Partial Repetition Rejection	96.8%	3-frame debounce & hysteresis state machine
Video Leakage Risk	Zero	100% on-device memory, vectors only

Test methodology, scope footnotes and the four full measurement tables · Read the engineering wiki →

What you get out of it

Once deployed, these interfaces deliver all metrics and visualization.

Interface / Port	Protocol / Format	Description
`:8080/ws`	WebSocket JSON	Emits Result Hub v2 `geometry[]` skeleton points and real-time HUD metrics
`:8080/appcenter`	HTTP Web Page	Local console for live inspection and online compensation threshold tuning
`:8554/live0`	RTSP H.264	Camera RTSP stream with hardware OSD burned-in

The WebSocket JSON contains only 33 coordinate pairs and joint angles, carrying zero video bytes, using under 20 KB/s bandwidth.

Every payload field and the Home Assistant discovery entities · Read the engineering wiki →

Porting It to Your Own System

The reusable asset is the pipeline 33-Keypoint Pose → Joint ROM Kinematics → Hysteresis State Machine → Compensation Detector.

Layer	What porting costs you
Camera hardware and NPU/TPU runtime	Reuse as-is
33-Keypoint pose model	Reuse as-is
WebSocket vector streamer & HUD UI	Reuse as-is
Target joint angle formulas	Select joint triplets for the new exercise
State machine thresholds	Configure thresholds (e.g. valgus angle, shrug angle)

How to deploy

Check the setup and its product list, then follow the deployment guide.

Hardware Sizing & Deployment Setup

Select site conditions to generate recommended bundles and guides, or let an AI assistant estimate hardware and deploy for you.

Self-service Consult & Deploy

Deployment Options

reCamera Pro

Deploy on reCamera Pro with RKNN INT8 model acceleration, full 33-point skeleton overlay, and interactive HUD dashboard.

2 steps · ~5min

Deployment Guide →

AI Camera

recamera_pro

We Are Glad to Be Your Hardware Partner !