AI Vehicle Damage Detection Using 3D Panel Geometry

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current systems for vehicle damage detection and estimation face challenges such as low accuracy, high computational cost, and inability to determine absolute damage sizes, limiting their utility in providing reliable repair estimates.

Innovation Solution

The implementation of a machine learning-based AI system that identifies vehicle damage by analyzing images, estimating the orientation and severity of impact, and requesting additional images when necessary, while using deep learning and homography to calculate absolute damage sizes, thereby improving accuracy and reducing computational costs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If Mask R-CNN or similar deep learning techniques are used for vehicle damage detection, then automation extent is improved, but manufacturing precision (damage detection accuracy) deteriorates due to low mAP values below 0.500

Engineering Contradiction:

Improveautomated damage detectionVSAvoiddamage detection accuracy

Core Design Contradiction:

Extent of automationVSManufacturing precision

Solution Approach 1:

The system segments the damage detection task into multiple specialized models: a vehicle detection model, a damage detection model, and a panel classification model. Each model focuses on specific aspects of damage identification, improving overall accuracy while maintaining automation. The segmentation of detection tasks allows each component to specialize and achieve higher precision in its specific function.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system transitions from 2D image analysis to 3D damage assessment by incorporating vehicle panel geometry and spatial relationships. The panel classification model uses 3D vehicle models to understand the spatial context of detected damage, enabling more accurate damage characterization beyond simple 2D image recognition.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Extent of automation

If complex deep learning models like Mask R-CNN are deployed, then extent of automation is improved, but use of energy (computational cost) worsens

Engineering Contradiction:

Improveautomated damage assessmentVSAvoidcomputational processing capacity

Core Design Contradiction:

Extent of automationVSUse of energy by moving object

Solution Approach 1:

The system divides the computational workload into separate, specialized models that can be executed independently and in parallel. The vehicle detection model, damage detection model, and panel classification model each process specific aspects of the input images, reducing the computational burden on any single model and enabling more efficient resource utilization.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system uses lightweight detection models that can be deployed on mobile devices or edge computing platforms, copying the essential detection functionality from complex server-based systems. This allows automated damage assessment to run locally with minimal computational overhead while maintaining acceptable accuracy levels.

Inventive Principle:

Principle #26Copying

3Ease of manufacture

If standard object detection models are used, then ease of manufacture is improved, but measurement precision (absolute damage size determination) deteriorates

Engineering Contradiction:

Improvesystem implementationVSAvoidabsolute damage area size

Core Design Contradiction:

Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system introduces vehicle panel 3D models as an intermediary between 2D damage detection and absolute size measurement. The panel classification model uses these 3D models to map detected damage regions to their corresponding physical panels, enabling accurate calculation of absolute damage areas by leveraging known panel geometries and dimensions.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system transforms the measurement parameter from relative damage area (percentage of image) to absolute damage area (physical dimensions) by incorporating vehicle-specific geometric parameters. The panel classification model uses vehicle make, model, and year information to retrieve accurate panel dimensions and calculate real-world damage sizes from detected damage regions.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11915479B1Systems and methods for automatedly identifying, documenting and reporting vehicle damage

Publication Date: 2024.02.27 SCOPE TECH HLDG

AI summary

Using a combination of image recognition, deep learning, and expert rules, disclosed systems and methods automatically identify, assess and estimate damage to vehicles based on input vehicle images. Analysis includes determinations as to type of damage, extent of damage and absolute size of damage. Disclosed embodiments include techniques of homography, allometry and semantic segmentation combined with more conventional object recognition technologies such as R-CNN in order to provide greater precision, more detailed analysis at lower computational cost.