mdashraf9723/body3d-realsense-measurements

Body3D — 3D Body Measurement from Intel RealSense D455

Automated body measurement extraction from .bag files recorded with the Intel RealSense D455 depth camera. Captures from 3 viewpoints (front, left, right) are fused into a unified 3D body scan, from which anthropometric measurements are computed.

📏 Measurements Extracted

Measurement	Method	Typical Accuracy
Height	Point cloud bounding box + landmark validation	< 1% error
Weight	Body volume integration × density (985 kg/m³)	~5-10% error
Neck Circumference	Cross-section slice + radial filtering	~10-15% error
Chest Circumference	Cross-section slice with torso isolation	~5-10% error
Waist Circumference	Narrowest torso cross-section	~5-10% error
Hip Circumference	Widest cross-section at hip level	~5% error
Wrist Circumference	Forearm-perpendicular cross-section	< 2% error
Shoulder Width	Bi-acromial landmark distance	< 1% error
BMI	Derived from height + estimated weight	—

---

🚀 How to Run It — Step by Step

Step 1: Install Dependencies

# Clone the repo
git clone https://huggingface.co/mdashraf9723/body3d-realsense-measurements
cd body3d-realsense-measurements

# Install Python packages
pip install -r requirements.txt

# On Ubuntu/Debian, you also need:
sudo apt-get install libgl1-mesa-glx libglib2.0-0

Headless server? Use pip install open3d-cpu instead of open3d.

---

Step 2: Record Your .bag Files (if not done already)

Set up your D455 camera at 3 positions around the person:

         [LEFT CAM]        
            ↓  (90°)       
                            
            🧍 ← Person (arms slightly out, A-pose)
                            
[FRONT CAM] →    ← [RIGHT CAM]
   (0°)              (-90°)    

Option A — Using Intel RealSense Viewer:

1. Open RealSense Viewer
2. Enable Depth + Color streams (1280×720, 30fps)
3. Click Record → save as front.bag
4. Move camera to left side → record left.bag
5. Move camera to right side → record right.bag

Option B — Using Python script:

import pyrealsense2 as rs
import time

pipeline = rs.pipeline()
config = rs.config()
config.enable_stream(rs.stream.depth, 1280, 720, rs.format.z16, 30)
config.enable_stream(rs.stream.color, 1280, 720, rs.format.bgr8, 30)
config.enable_record_to_file("front.bag")  # change name for each view

pipeline.start(config)
time.sleep(5)  # Record 5 seconds
pipeline.stop()
print("Done!")

Recording Tips:

• Stand 1.5–2.5m from camera
• Person should stand still, arms slightly away from body (A-pose)
• Record 3–5 seconds per view
• Good lighting, avoid direct sunlight (interferes with IR depth sensor)
• Clear background helps with person segmentation

---

Step 3: Run the Measurement Pipeline

Put your .bag files in the same folder as the scripts, then run:

Multi-View (3 .bag files — BEST accuracy):

python pipeline.py \
    --front front.bag \
    --left left.bag \
    --right right.bag \
    --output my_measurements.json

Multi-View with known camera angles:

python pipeline.py \
    --front front.bag \
    --left left.bag \
    --right right.bag \
    --angles 0 -90 90 \
    --output my_measurements.json

Single View (1 .bag file — less accurate):

python pipeline.py --bag front.bag --output my_measurements.json

Test/Demo mode (no camera needed):

python pipeline.py --demo

---

Step 4: Read the Results

The output my_measurements.json will look like:

{
  "height_cm": 175.0,
  "estimated_weight_kg": 72.5,
  "bmi": 23.7,
  "neck_circumference_cm": 38.2,
  "chest_circumference_cm": 94.5,
  "waist_circumference_cm": 82.1,
  "hip_circumference_cm": 96.3,
  "wrist_circumference_cm": 16.4,
  "shoulder_width_cm": 44.0
}

It also prints a formatted table in the terminal:

============================================================
  BODY MEASUREMENTS
============================================================
  Height............................. 175.0 cm
  Estimated Weight................... 72.5 kg
  BMI................................ 23.7
  ----------------------------------------
  Neck Circumference................. 38.2 cm
  Chest Circumference................ 94.5 cm
  Waist Circumference................ 82.1 cm
  Hip Circumference.................. 96.3 cm
  Wrist Circumference................ 16.4 cm
  ----------------------------------------
  Shoulder Width..................... 44.0 cm
============================================================

---

Quick Reference

What you have	Command
3 .bag files (front, left, right)	python pipeline.py --front front.bag --left left.bag --right right.bag
1 .bag file	python pipeline.py --bag recording.bag
No camera, just testing	python pipeline.py --demo

All CLI Options

python pipeline.py --help

Options:
  --demo                Run with synthetic test data (no camera needed)
  --bag FILE            Single .bag file path
  --front FILE          Front view .bag file
  --left FILE           Left view .bag file
  --right FILE          Right view .bag file
  --angles FLOAT...     Camera angles in degrees (e.g. 0 -90 90)
  --output FILE         Output JSON path (default: measurements.json)
  --voxel-size FLOAT    Point cloud resolution in meters (default: 0.01)
  --max-depth FLOAT     Max depth from camera in meters (default: 3.0)
  --save-ply FILE       Save merged point cloud to PLY file

---

🏗️ Pipeline Architecture

.bag files (3 views) → bag_reader.py → registration.py → landmarks.py → measurements.py → JSON output

Modules:

1. bag_reader.py — Reads RealSense .bag files via pyrealsense2, applies spatial/temporal/hole-filling filters, extracts aligned depth+color frames, generates Open3D point clouds with camera intrinsics
2. registration.py — FPFH feature extraction + RANSAC global registration + Point-to-plane ICP refinement to merge 3 views; DBSCAN clustering for person segmentation; RANSAC floor plane removal
3. landmarks.py — MediaPipe Pose detection (33 body landmarks), 2D→3D lifting via depth unprojection with robust neighborhood depth sampling
4. measurements.py — Angular-sweep boundary extraction on cross-sectional slices with radial distance filtering to isolate torso from arms; volume integration for weight; multi-method weighted estimation
5. pipeline.py — End-to-end CLI orchestrator with single-view, multi-view, and demo modes

📁 File Structure

body3d-realsense-measurements/
├── pipeline.py          # Main entry point (CLI)
├── bag_reader.py        # Module 1: RealSense .bag file reader
├── registration.py      # Module 2: Multi-view point cloud registration
├── landmarks.py         # Module 3: MediaPipe body landmarks
├── measurements.py      # Module 4: Body measurement computation
├── requirements.txt     # Python dependencies
└── README.md            # This file

---

🧪 How Measurements Work

Cross-Section Circumference (Angular Sweep Method)

For each body measurement (chest, waist, hip, neck, wrist):

1. Slice the point cloud at the landmark height (±1-2.5cm tolerance)
2. Filter by radial distance from body center (excludes arms for torso measurements)
3. Project slice points to 2D plane
4. Angular sweep: Cast 72-90 rays from centroid, find outermost point per angle bin
5. Compute perimeter of the ordered boundary polygon

Weight Estimation

Three methods combined with weighted average:

1. Volume integration (weight: 50%): Integrate cross-sectional areas along height axis, multiply by body density (985 kg/m³)
2. Measurement regression (weight: 15%): Empirical formula from height + chest + waist + hip
3. Hamwi formula (weight: 5%): Clinical height-based estimate

Person Segmentation

1. RANSAC floor detection: Fits plane to largest flat surface, removes floor points
2. DBSCAN clustering: Groups remaining points by proximity (5cm neighborhood)
3. Largest cluster: Selected as the person

---

🧪 Validated Results (Synthetic)

Tested on synthetic body with known ground truth dimensions:

✓ height_cm.............. GT: 175.0  Est: 175.0  Err: 0.0 (0.0%)
✓ chest_circumference.... GT:  94.2  Est:  85.7  Err: 8.5 (9.1%)
✓ waist_circumference.... GT:  81.7  Est:  85.6  Err: 3.9 (4.7%)
✓ hip_circumference...... GT:  97.4  Est:  93.1  Err: 4.3 (4.5%)
⚠ neck_circumference..... GT:  37.7  Est:  43.0  Err: 5.3 (14.2%)
✓ wrist_circumference.... GT:  15.7  Est:  15.6  Err: 0.1 (0.7%)
✓ shoulder_width......... GT:  44.0  Est:  44.0  Err: 0.0 (0.0%)

---

🔧 Advanced: SMPL-Based Measurements

For higher accuracy (~1-2cm error), you can integrate SMPL body model fitting:

1. Register at smpl.is.tue.mpg.de and download model files
2. Install smplx: pip install smplx
3. Install SMPL-Anthropometry: pip install git+https://github.com/DavidBoja/SMPL-Anthropometry

The pipeline can be extended to fit SMPL to the merged point cloud and extract measurements from the parametric mesh, giving more precise circumferences through anatomical priors.

---

📚 References

• ArtEq (CVPR 2023) — SE(3)-equivariant SMPL fitting from point clouds
• ETCH (2025) — Clothed human body estimation from point clouds
• Pose-Independent Anthropometry (2025) — Body measurements from sparse landmarks
• A2B (2024) — Bidirectional anthropometric ↔ SMPL-X shape mapping

---

⚠️ Limitations

• 3-view gaps: Back of torso may have incomplete coverage → affects waist/hip accuracy
• Clothing: Loose clothing adds to circumferences. Use ETCH method for clothed subjects
• Weight estimation: Volume-based weight has ~5-10% error; depends on scan completeness
• Small features: Wrist/neck require good point cloud density at those locations
• Coordinate system: Assumes Y-up. Different camera orientations may need axis remapping