Vehicle inspection system and method with adaptive template matching and temporal defect tracking

1 . A vehicle inspection system comprising:

an imaging assembly disposed around a vehicle inspection lane, the imaging assembly including a plurality of cameras positioned to capture multi-view images of at least a portion of a vehicle;

a memory storing instructions;

one or more processors configured for executing the instructions to:

retrieve a vehicle identifier of the vehicle;

determine, from the multi-view images, a spatial pose of the vehicle relative to the imaging assembly;

retrieve, from a template library, a reference template corresponding to the vehicle identifier;

compute a geometric transformation that aliens the reference template to the determined spatial pose;

compute, between the aligned reference template and the captured images, one or more spatially-registered similarity maps;

generate a deviation map from the one or more similarity maps;

compare the deviation map against deviation maps from prior inspections of the same vehicle to distinguish persistent anomalies from transient conditions; and

generate an inspection report identifying classified vehicle defects with spatial locations and severity scores,

wherein the imaging assembly includes polarization-controlled imaging sensors, and wherein the one or more processors are configured to:

separate diffuse and specular reflection components from the captured images,

compute a first similarity map on the diffuse reflection components to detect color and texture anomalies, and

compute a second similarity map on the specular reflection patterns to detect surface geometry deviations.

2 . The system of claim 1 , wherein the one or more processors are further configured to apply photometric normalization to at least one of the reference template and the captured images to compensate for illumination differences.

3 . The system of claim 1 , wherein the one or more spatially-registered similarity maps include two or more of:

a deep-feature correlation map derived from convolutional neural network features,

an edge-distance map computed from distance transforms of edge maps, and

a photometrically-normalized intensity correlation map.

4 . The system of claim 3 , wherein the one or more spatially-registered similarity maps further include a phase-only correlation map computed to capture structural similarity independent of intensity variations.

5 . The system of claim 1 , wherein the reference template is a per-vehicle template associated with the vehicle identifier, and wherein the one or more processors are further configured to update the per-vehicle template by incorporating regions confirmed as normal and excluding regions confirmed as defective.

6 . The system of claim 5 , wherein the one or more processors are configured to initialize the per-vehicle template at a first inspection visit and evolve the per-vehicle template across subsequent inspections using gated learning based on inspection outcomes.

7 . The system of claim 1 , wherein the template library includes both:

a per-vehicle template specific to the vehicle identifier, and

a class template derived from fleet-level statistical aggregation of multiple vehicles of a same class, and

wherein the one or more processors weight evidence from the per-vehicle template and the class template using learned uncertainty weights.

8 . The system of claim 7 , wherein the one or more processors apply region-specific adaptive thresholds to the deviation map based on fleet-level variance priors that encode expected variability for corresponding regions of the vehicle class.

9 . The system of claim 8 , wherein the fleet-level variance priors include seasonal priors that modulate expected variance based on time of year and geographic region to account for weather-related conditions.

10 . The system of claim 1 , wherein retrieving the reference template comprises synthesizing the reference template by:

accessing a three-dimensional vehicle model corresponding to the vehicle identifier,

rendering the three-dimensional vehicle model under illumination conditions estimated from the captured images, and

projecting the rendered model to viewpoints corresponding to the plurality of cameras.

11 . The system of claim 1 , wherein the plurality of cameras include undercarriage imaging sensors, and wherein the one or more processors are configured to:

register sequential undercarriage frames captured as the vehicle moves through the inspection lane,

construct an undercarriage panorama in a vehicle-centric coordinate frame, and

detect at least one of missing components, foreign objects, loose components, and fluid leaks by comparing the undercarriage panorama to an undercarriage reference template.

12 . The system of claim 1 , wherein the one or more processors are further configured to:

receive one or more exemplar image patches depicting a target defect type,

compute exemplar-conditioned correlation between features of the exemplar image patches and features of the captured images, and

perform bounding-box regression conditioned on the exemplar-conditioned correlation to detect instances of the target defect type without retraining a detection model.

13 . The system of claim 12 , wherein the one or more exemplar image patches are automatically mined from confirmed defects in prior inspections of a fleet of vehicles.

14 . The system of claim 1 , wherein computing the geometric transformation includes:

computing a coarse translational alignment using phase-only correlation,

refining the coarse alignment to an affine transformation using enhanced correlation coefficient optimization, and

applying a local non-rigid refinement to accommodate vehicle deformations.

15 . The system of claim 1 , wherein the imaging assembly further includes at least one depth sensor, and wherein the one or more processors are configured to:

back-project candidate defect regions to three-dimensional space using depth data and camera calibration parameters, and

require corroboration of a defect from multiple camera viewpoints based on epipolar geometry before classifying as a confirmed defect.

16 . The system of claim 15 , wherein the one or more processors compute severity scores based on at least one of:

depth or displacement magnitude measured from the at least one depth sensor, affected surface area,

temporal persistence across multiple inspections, and

deviation magnitude in the deviation map.

17 . A method for automated vehicle inspection, comprising:

capturing, by an imaging assembly having polarization-controlled imaging sensors, multi-view images of a vehicle;

retrieving a vehicle identifier and determining a spatial pose of the vehicle;

retrieving, from a template library, a reference template corresponding to the vehicle identifier;

aligning the reference template to the determined spatial pose by computing a geometric transformation;

computing, between the aligned reference template and the captured images, a plurality of spatially-registered similarity maps including at least a deep-feature correlation map and an edge-distance map;

separating diffuse and specular reflection components from the captured images;

computing a first similarity map on the diffuse reflection components to detect color and texture anomalies;

computing a second similarity map on the specular reflection patterns to detect surface geometry deviations

generating a deviation map by fusing the plurality of similarity maps;

comparing the deviation map against prior deviation maps from previous inspections of the vehicle to identify persistent anomalies; and

generating an inspection report with classified defects, spatial locations, and severity scores.

18 . The method of claim 17 , further comprising:

updating the reference template based on confirmed inspection outcomes by:

incorporating regions confirmed as defect-free into the reference template,

excluding regions confirmed as defective from the reference template, and

evolving the reference template across multiple inspections of the vehicle using gated learning;

wherein the reference template is a per-vehicle template specific to the vehicle identifier.

19 . The method of claim 17 , further comprising:

receiving one or more exemplar patches depicting a specific defect type;

computing spatial correlation between features extracted from the exemplar patches and features extracted from the captured images; and

localizing additional instances of the specific defect type using template-conditioned regression based on the spatial correlation, thereby enabling detection of new defect types without retraining a detection model.