Scene calibration configurations
Files and source keys
| Configuration | Source keys | Master source |
|---|---|---|
WS1_2.json |
video1-video6 | video4 |
WS3.json |
video1-video3 | video2 |
WS4_5.json |
video1-video4 | video1 |
WS6_7.json |
video1-video3 | video2 |
Use the scene JSON and the --source videoN value associated with the raw CSV in configs/source_manifest.csv. Pixel centers and bounding-box corners must use the same image coordinate view as the control points. Source-image coordinates are mapped into the scene's master-image frame before metric conversion.
Configuration fields
| Field | Meaning and units |
|---|---|
template_only |
false for the supplied actual scene configurations |
scene_id |
Observation-group identifier |
master_source |
Source key defining the master-image coordinate frame |
origin_px |
Two-element [x,y] origin in master-image pixels |
rotation_rad |
Zero for the supplied road-aligned workflow |
reference_segments |
Measured segments expressed in master-image pixels and meters |
reference_segments[].point_a, point_b |
Two-element pixel endpoints in the master image |
reference_segments[].real_distance_m |
Positive measured length in meters |
sources.videoN.src_points |
Nx2 source-image fitting control points, in pixels |
sources.videoN.master_points |
Matched Nx2 master-image fitting control points, in pixels |
The master source has an empty source entry because its image frame is already the master frame. Non-master fitting arrays must have equal lengths, at least four point pairs and noncollinear finite coordinates.
Transformation
For each measured reference segment, calculate its real length divided by its pixel length. The metric scale is the arithmetic mean of these ratios, in meters per pixel.
For a non-master source, coordinate_transform.py estimates the source-to-master homography from the matched fitting points using RANSAC with a 3-pixel threshold and RNG seed 0. For a master source, the source-to-master transform is the identity. After mapping a pixel point to master coordinates (u,v), the output is:
x = (u - origin_px[0]) * scale
y = (origin_px[1] - v) * scale
Image coordinates follow the pixel convention, with y increasing downwards. Metric x follows the aligned road axis and metric y increases upwards. Both output coordinates are in meters. Vehicle centers and all four bounding-box corners use this same transformation.
Usage
python coordinate_transform.py --input data/raw/WS3_raw_detection_data_2.csv --calibration calibrations/WS3.json --source video2 --output runs/ws3_video2_metric.csv
python estimate_vehicle_length.py --input data/raw/WS3_raw_detection_data_2.csv --calibration calibrations/WS3.json --source video2 --output runs/ws3_video2_lengths.csv
Malformed endpoints, nonpositive reference distances, unmatched point arrays, degenerate homographies and points on a homography horizon produce explicit errors. The supplied scene JSONs provide the fitting values used by the processing commands.