🎓 Lesson 7
D4
Sub-Pixel Edge Detection Accuracy vs. Lighting Conditions
Sub-pixel edge detection makes robot vision see object edges more precisely than a single camera pixel—especially when lighting changes.
🎯 Learning Objectives
- ✓ Analyze how illumination intensity and angular direction affect sub-pixel edge localization error (in pixels) using empirical calibration data
- ✓ Design a lighting configuration (diffuse vs. directional, color temperature, intensity range) to achieve ≤0.15 px edge localization uncertainty for granite surface imaging
- ✓ Calculate the expected sub-pixel edge error given measured SNR, edge contrast ratio, and PSF width using the Cramér-Rao lower bound approximation
- ✓ Explain the trade-off between lighting-induced specular reflection and edge detectability in high-albedo rock surfaces
- ✓ Apply ISO/IEC 19794-6:2022 guidelines to validate edge detection accuracy under variable mine-site lighting conditions
📖 Why This Matters
In autonomous drilling and blast-hole verification systems, robots must locate hole centers or fracture boundaries within ±0.2 mm at 2 m working distance—a requirement demanding <0.1 px edge localization accuracy. Poor lighting causes shadows, glare, or low contrast on dusty, wet, or oxidized rock faces, degrading sub-pixel edge detection by up to 3×. This lesson bridges computer vision theory with real mining site constraints—so your algorithms don’t fail at noon or in fog.
📘 Core Principles
Sub-pixel edge detection relies on modeling the 1D intensity profile across an edge (e.g., as a step, ramp, or hyperbolic tangent) and estimating its true zero-crossing or inflection point via interpolation or optimization. Accuracy depends on three interdependent factors: (1) Signal-to-Noise Ratio (SNR), governed by illumination intensity and sensor gain; (2) Edge contrast (ΔI/Iₘₑₐₙ), reduced by diffuse scattering or ambient light contamination; and (3) Point Spread Function (PSF) width, which blurs edges and couples with lighting geometry. Crucially, non-uniform lighting introduces bias—not just noise—by distorting local gradients, violating the constant-background assumption in most sub-pixel estimators (e.g., Zernike moment or Gaussian derivative methods).
📐 Cramér-Rao Lower Bound for Edge Localization Error
The theoretical minimum variance (σ²) of any unbiased edge position estimator depends on image gradient strength and noise—directly modulated by lighting. This bound quantifies the best possible accuracy under given optical conditions.
Cramér-Rao Lower Bound (CRLB) for Edge Position
σₓ² = 1 / (2 × SNR × G²)Theoretical lower limit on variance of unbiased edge position estimator; G is normalized gradient magnitude per pixel.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| σₓ | Standard deviation of edge position estimate | px | Localization uncertainty along edge normal direction |
| SNR | Signal-to-noise ratio | dimensionless (linear) | Ratio of edge signal power to noise power in ROI |
| G | Normalized edge gradient magnitude | px⁻¹ | First derivative of intensity profile, scaled to [0,1] per pixel |
Typical Ranges:
Optimized LED lighting on clean granite: 0.03 – 0.06 px
Overcast daylight on weathered sandstone: 0.10 – 0.25 px
💡 Worked Example
Problem: A monochrome camera (pixel pitch = 3.45 µm) images a granite surface under LED lighting (6500 K, irradiance = 1200 lux). Measured edge SNR = 28 dB, average edge contrast ratio ΔI/Iₘₑₐₙ = 0.62, and PSF FWHM = 1.8 pixels. Calculate theoretical minimum edge localization standard deviation σₓ in pixels.
1.
Step 1: Convert SNR from dB → linear: SNR_lin = 10^(28/10) ≈ 631
2.
Step 2: Compute effective gradient magnitude G ≈ (π/2) × (ΔI/Iₘₑₐₙ) / (PSF_FWHM × √(2 ln 2)) ≈ (1.57)(0.62) / (1.8 × 0.833) ≈ 0.65 per pixel
3.
Step 3: Apply CRLB: σₓ² = 1 / (2 × SNR_lin × G²) = 1 / (2 × 631 × 0.65²) ≈ 1 / 532 ≈ 0.00188 → σₓ ≈ √0.00188 ≈ 0.043 px
Answer:
The theoretical minimum edge localization error is 0.043 px (≈ 0.15 µm), well within the 0.15 px target. However, real-world bias from lighting non-uniformity increases observed error to 0.11 px—confirming need for active lighting control.
🏗️ Real-World Application
At BHP’s Olympic Dam underground drill bay, a vision-guided robotic drill verifies blast-hole positions before charging. Under inconsistent HID overhead lighting (±40% irradiance variation), sub-pixel edge detection on hole rims showed 0.21 px RMS error—causing 0.7 mm misalignment at 2.1 m standoff. Engineers installed synchronized, ring-mounted 5000 K LEDs with diffusers and intensity feedback control (per ISO/IEC 19794-6 Annex D), reducing edge error to 0.09 px and cutting rework by 68% over 6 months.
🔧 Interactive Calculator
🔧 Open Industrial Robot Vision Guidance & Integration Calculator📋 Case Connection
📋 Vision-Guided Palletizing Robot for Mixed-SKU E-Commerce Fulfillment
Unstructured tote input with random part orientation, variable box dimensions, and reflective surfaces causing glare
📋 Vision-Guided Arc Welding Cell for Automotive Chassis Assembly
Thermal distortion-induced seam deviation (>1.8 mm) between stamped aluminum panels; inconsistent joint gap due to fixtu...
📋 3D Vision-Guided Bin-Picking for Aerospace Fastener Kits
Highly reflective titanium fasteners (M3–M8), nested geometry, dense packing, and strict traceability requirements
📋 Vision-Guided Deburring Robot for Cast Aluminum Engine Blocks
Variable burr height (0.1–2.4 mm) and location due to inconsistent casting; complex freeform surfaces limiting fixed-too...