3. Data acquisition and preprocessing for crash deformation

Geometric deep learning models for crash deformation are only as trustworthy as the displacement fields they are trained on. Before any graph neural network or neural operator sees a single vertex, the raw measurements from a crash test must be converted into a dense, topologically consistent, physically meaningful deformation field. This chapter walks through the four dominant acquisition modalities — stereo digital image correlation (DIC), photogrammetry, laser scanning, and the mesh registration pipelines that fuse them — and shows how each contributes to a usable deformation dataset.

What the downstream model actually needs

A deformation field for crash analysis is a mapping from a reference configuration X to a deformed configuration x, expressed as displacement u(X) = x − X, ideally sampled on a mesh whose connectivity is identical across all frames and all specimens. Three properties matter most:

• Correspondence: vertex i in the reference mesh must map to the same material point in every deformed state. Without this, gradients and strain are meaningless.
• Density and coverage: high-gradient regions (folds, tear zones, spot-weld neighborhoods) need finer sampling than flat panels.
• Metric accuracy: displacement errors should be well below the smallest strain of interest; for automotive sheet metal, sub-millimeter accuracy at 1–2 m standoff is a common target.

No single sensor satisfies all three. DIC gives dense surface displacement with excellent correspondence but only on visible faces; photogrammetry gives sparse but globally consistent 3D point tracks; laser scanning gives high-fidelity geometry but no temporal correspondence by itself. The pipeline is therefore a fusion problem.

Stereo DIC: dense surface displacement

Stereo DIC uses two synchronized high-speed cameras viewing a speckle-patterned surface. The workflow is well established but the details determine data quality:

1. Speckle application: a stochastic black-on-white pattern with speckle size roughly 3–5 pixels at the working distance. For a 1.5 m field of view with 2048×2048 sensors, that means speckles of about 2–4 mm.
2. Calibration: a dot-grid or checkerboard target is imaged in 15–25 poses covering the measurement volume. A typical reprojection RMS below 0.05 pixels indicates a usable calibration; above 0.1 pixels, re-calibrate.
3. Subset correlation: a subset size of 31–51 pixels with a step of 5–10 pixels balances spatial resolution against correlation robustness. In high-strain zones, smaller subsets track local gradients but risk decorrelation.
4. Triangulation and 3D reconstruction: matched points from both cameras are triangulated into a point cloud per frame.

The output is a time series of 3D point clouds on the visible surface. The critical preprocessing step is projecting these points onto a common reference mesh — the topic of the registration section below. A practical pitfall: DIC loses track in regions of severe out-of-plane rotation or occlusion (e.g., a folding fender occluding the door). Those gaps must be flagged, not interpolated blindly, because a GNN will happily learn from fabricated displacements.

Photogrammetry: global consistency and sparse tracks

Photogrammetry (structure-from-motion or bundle adjustment over coded targets) is typically used to establish the global reference frame and to track a modest number of retro-reflective markers bonded to the specimen. Its strength is that it does not decorrelate: coded targets remain identifiable through large rigid-body motion and moderate deformation. Its weakness is sparsity — tens to a few hundred points, not millions.

In practice, photogrammetry serves two roles. First, it defines the specimen coordinate system: three or more reference targets on a rigid fixture establish origin and axes, and all other modalities are registered into that frame. Second, marker displacements provide a low-frequency check on DIC — if the DIC-derived displacement at a marker location disagrees with the photogrammetric track by more than, say, 0.5 mm, something in the calibration or correlation is wrong.

Laser scanning: geometry ground truth

Terrestrial or arm-mounted laser scanners capture the post-crash geometry at high fidelity, often 0.1–0.5 mm point spacing. Scans are acquired before and after the event (or on an undeformed reference specimen), and the undeformed scan becomes the geometric backbone of the reference mesh. Scanning is slow and line-of-sight limited, so it is usually performed on the final deformed state rather than continuously.

The scanner's contribution is geometric truth: it resolves folds, thinning, and local buckling that DIC may smooth over. The cost is that scans carry no temporal information and must be registered to the DIC/photogrammetry frame using shared fiducial targets or geometric features.

The registration pipeline

Registration is where the modalities are fused into a single deformation field. A robust pipeline proceeds coarse-to-fine:

1. Global alignment (rigid, 6-DOF)
   - Match photogrammetry fiducials to scan/DIC fiducials
   - Solve Procrustes / Umeyama; check RMS residual < 0.3 mm
2. Reference mesh construction
   - Decimate undeformed scan to target vertex count (e.g., 50k–200k)
   - Preserve feature edges (tears, weld lines) via curvature-aware decimation
3. Per-frame non-rigid registration
   - Coherent Point Drift or non-rigid ICP from DIC cloud to reference mesh
   - Regularize with as-rigid-as-possible or bounded-strain prior
4. Correspondence transfer
   - Assign each reference vertex the nearest DIC track; flag gaps
5. Field assembly
   - u(X) = x_registered − X; store per-frame, per-vertex
   - Compute strain via Green-Lagrange tensor on the mesh

Two details deserve emphasis. First, non-rigid registration must be bijective enough to avoid fold-over: if the deformation gradient determinant J drops below zero anywhere, the mapping has inverted and the strain field is garbage. A practical guard is to reject frames where more than 0.1% of elements have J ≤ 0 and re-run with a stiffer regularization. Second, the reference mesh topology must be frozen before registration; changing vertex count between frames destroys correspondence and silently corrupts the dataset.

Putting it together: a data quality checklist

The figure below summarizes how the four modalities feed the registration pipeline and where quality gates sit. Each gate corresponds to a measurable criterion that must pass before the field enters the training set.

Treat the gates as hard filters, not suggestions. A dataset that passes all four gates on every frame is small but trustworthy; a dataset that skips Gate 3 will train a model that predicts inverted elements and unphysical strain concentrations. In the next stage, these fields become the node features and edge attributes of the deformation graph — but that only works if the correspondence and metric accuracy established here are sound.