Skip to main content

​
Introduction

The EVM Vital Signs Monitor is a contactless vital signs monitoring system that uses Eulerian Video Magnification (EVM) to extract heart rate (HR) and respiratory rate (RR) from standard video input. The system processes video in real-time, detecting subtle color changes in the face region that correspond to physiological signals.
The system achieves 50-60% faster performance than traditional dual-pass EVM implementations through a single-pass dual-band processing optimization.

​
System Workflow

The complete processing pipeline consists of five main stages:

​
Processing Stages

The system first detects faces in the video stream and extracts a stable Region of Interest (ROI):
  • Multiple detector backends: Haar Cascade, MTCNN, YOLO, MediaPipe
  • Temporal stabilization: Weighted averaging over last 5 frames reduces jitter
  • ROI weights: [0.1, 0.15, 0.2, 0.25, 0.3] (recent frames weighted higher)
  • Adaptive boundaries: Ensures ROI stays within frame limits
Source: src/face_detector/manager.py:95-130
Each ROI frame is decomposed into a Laplacian pyramid for multi-resolution analysis:
  • Gaussian pyramid: Successive downsampling creates multiple resolution levels
  • Laplacian pyramid: Difference between Gaussian levels captures spatial frequencies
  • Default levels: 3 levels optimized for Raspberry Pi performance
  • Dual-band optimization: Single pyramid construction supports both HR and RR extraction
Different pyramid levels capture different spatial frequencies:
  • Level 3: Higher spatial frequency (optimal for heart rate)
  • Level 2: Lower spatial frequency (optimal for respiration)
Source: src/evm/pyramid_processing.py:98-126
Separate bandpass filters isolate physiological frequency bands:Heart Rate Band (0.8-3 Hz):
  • Frequency range: 48-180 BPM
  • Applied to pyramid level 3
  • Butterworth bandpass filter (order 2)
Respiratory Rate Band (0.2-0.8 Hz):
  • Frequency range: 12-48 breaths per minute
  • Applied to pyramid level 2
  • Independent filtering allows simultaneous extraction
Source: src/evm/temporal_filtering.py:30-60
Filtered signals are amplified to enhance subtle variations:
  • Heart rate amplification: α = 30
  • Respiratory rate amplification: α = 50
  • Spatial averaging: Green channel for HR, all channels for RR
The amplification factors are carefully tuned to maximize signal visibility while avoiding motion artifacts.Source: src/evm/evm_core.py:98-108
FFT-based spectral analysis extracts dominant frequencies:
  1. Signal preprocessing: Detrending and normalization
  2. Hamming window: Reduces spectral leakage
  3. FFT computation: Converts to frequency domain
  4. Peak detection: Finds dominant frequency in physiological range
  5. Validation: Ensures result falls within plausible bounds
Source: src/evm/signal_analysis.py:127-158

​
Key Architectural Components

​
Core Modules

ModuleLocationResponsibility
EVMProcessorsrc/evm/evm_core.pyOrchestrates dual-band EVM processing
FaceDetectorsrc/face_detector/manager.pyManages face detection and ROI stabilization
Pyramid Processingsrc/evm/pyramid_processing.pyBuilds and manipulates Laplacian pyramids
Temporal Filteringsrc/evm/temporal_filtering.pyApplies bandpass filters to isolate frequencies
Signal Analysissrc/evm/signal_analysis.pyFFT-based frequency extraction
Configurationsrc/config.pyCentral parameter definitions

​
Data Flow

1

Video Frame Acquisition

Input frames are captured at 30 FPS (configurable) and resized to 320x240 for processing efficiency.
2

ROI Extraction

Face detector identifies face region and applies temporal smoothing to maintain stable ROI across frames.
3

Single-Pass Pyramid Construction

One Laplacian pyramid stack is built for the entire video buffer (typically 30+ frames).
4

Dual-Band Processing

Two pyramid levels are extracted and filtered independently:
  • Level 3 tensor → HR bandpass → α=30 amplification → Green channel signal
  • Level 2 tensor → RR bandpass → α=50 amplification → All-channel signal
5

Frequency Extraction

Both signals undergo FFT analysis to extract dominant frequencies within physiological ranges.

​
Performance Optimization

Key Innovation: The dual-band processing approach builds Laplacian pyramids only once, then applies different temporal filters to different pyramid levels. This eliminates the need for separate pyramid construction for HR and RR, achieving 50-60% performance improvement.

​
Traditional vs. Optimized Approach

Traditional (Two-Pass): 
# First pass: Heart rate
pyramids_hr = build_pyramids(frames)
filtered_hr = bandpass(pyramids_hr, 0.8, 3.0)
hr_result = analyze(filtered_hr)

# Second pass: Respiratory rate (rebuilds pyramids!)
pyramids_rr = build_pyramids(frames)
filtered_rr = bandpass(pyramids_rr, 0.2, 0.8)
rr_result = analyze(filtered_rr)
Optimized (Single-Pass): 
# Single pyramid construction
pyramids = build_pyramids(frames)

# Extract different levels
tensor_hr = extract_level(pyramids, level=3)
tensor_rr = extract_level(pyramids, level=2)

# Parallel filtering and analysis
hr_result = analyze(bandpass(tensor_hr, 0.8, 3.0))
rr_result = analyze(bandpass(tensor_rr, 0.2, 0.8))
Reference: src/evm/evm_core.py:40-109

​
Configuration Parameters

Key system parameters are defined in src/config.py:1-33: 
# Heart Rate Parameters
ALPHA_HR = 30          # Amplification factor
LOW_HEART = 0.83       # Low cutoff (Hz)
HIGH_HEART = 3.0       # High cutoff (Hz)
MIN_HEART_BPM = 40     # Validation range
MAX_HEART_BPM = 250

# Respiratory Rate Parameters
ALPHA_RR = 50          # Amplification factor
LOW_RESP = 0.18        # Low cutoff (Hz)
HIGH_RESP = 0.5        # High cutoff (Hz)
MIN_RESP_BPM = 8       # Validation range
MAX_RESP_BPM = 35

# Processing Parameters
FPS = 30               # Frame rate
LEVELS_RPI = 3         # Pyramid levels
TARGET_ROI_SIZE = (320, 240)  # ROI dimensions

​
Next Steps

Eulerian Video Magnification

Deep dive into the EVM technique and dual-band implementation

Signal Processing

Learn about temporal filtering and FFT-based frequency analysis

Face Detection

Explore ROI extraction and stabilization techniques

API Reference

Browse the complete API documentation

Build docs developers (and LLMs) love

Get started for freeTalk to us