avm/docs/TECHNICAL_SPECS.md

Technical Specifications

Camera Parameters

Fisheye Camera Specifications

• Focal Length: 910.0 pixels
• Pixel Size: 3.0μm × 3.0μm
• Image Resolution: 1280×960 pixels
• Fisheye Scaling Factor: 0.5
• Undistortion Scaling: 1.55

Distortion Model

The system uses a polynomial fisheye distortion model:

θ_distorted = θ_undistorted + k₁θ³ + k₂θ⁵ + k₃θ⁷ + k₄θ⁹

Default distortion coefficients:

• k₁ = -0.05611147
• k₂ = -0.05377447
• k₃ = 0.0115717
• k₄ = 0.0030788

Calibration Board Specifications

Pattern Configuration

• Pattern Type: Rectangular grid
• Grid Size: 2×4 (2 rows, 4 columns)
• Corner Count: 8 corners total
• Detection Method: Histogram-based adaptive thresholding

Detection Parameters

• Contrast Enhancement: 2.5× gamma correction
• Valid Detection Region: Y-axis 20%-70% of image height
• Corner Refinement: Sub-pixel accuracy using cornerSubPix
• Backup Detection: Adaptive threshold with 401×401 kernel

Bird's Eye View Parameters

Image Dimensions

• Front/Back Views: 792×305 pixels
• Left/Right Views: 1131×281 pixels
• Final Stitched Image: Variable based on layout

Transformation Matrices

Each camera view uses homography transformation calculated from:

• Source: Detected corner points in undistorted image
• Target: Predefined bird's eye view coordinates

Rotation Corrections

• Front View: 0° (no rotation)
• Back View: 180°
• Left View: 90° clockwise
• Right View: 90° counter-clockwise

Image Stitching Parameters

Layout Configuration

• Back Image Offset: Y = 643 pixels
• Right Image Offset: X = 398 pixels
• Vehicle Overlay: Center positioned with alpha blending

Blending Masks

• Individual masks for each camera view
• Smooth transition zones between adjacent views
• Alpha channel support for vehicle model overlay

Performance Characteristics

Processing Pipeline Timing

1. Image Loading: ~10ms per image
2. Undistortion: ~50ms per image
3. Corner Detection: ~100-200ms per image
4. Perspective Transform: ~30ms per image
5. Image Stitching: ~100ms total

Memory Requirements

• Input Images: ~5MB (4 × 1280×960×3)
• Intermediate Results: ~20MB
• Output Image: ~2-3MB
• Total Peak Usage: ~30-40MB

Algorithm Details

Corner Detection Algorithm

1. Preprocessing:

   • Convert to grayscale
   • Apply gamma correction for contrast enhancement
   • Set valid detection region (20%-70% Y-axis)

2. Binarization:

   • Histogram analysis for optimal threshold
   • Bimodal distribution detection
   • Adaptive fallback if histogram method fails

3. Contour Analysis:

   • Find contours in binary image
   • Filter by area (minimum threshold)
   • Approximate polygonal shapes

4. Rectangle Validation:

   • Verify 4-sided polygon structure
   • Check geometric constraints
   • Sort corners in consistent order

5. Coordinate Transformation:

   • Transform from fisheye to undistorted coordinates
   • Apply sub-pixel refinement
   • Validate final corner positions

Homography Calculation

Uses OpenCV's findHomography with:

• Method: Direct Linear Transform (DLT)
• Source Points: 8 detected corners
• Target Points: Predefined bird's eye coordinates
• Robustness: No RANSAC (assumes clean input)