Vehicle Damage Detection from Gate Camera Imagery – Oriented Bounding Boxes (Case Study)
Location: (Gate camera)
Project Overview
Nazru’s AI platform was deployed to automatically detect and localise vehicle damage using high‑resolution images from gate cameras. Each damage instance is output as a rotating rectangle (oriented bounding box) tightly fitted to the damage orientation. The system classifies damages into 11 distinct types, enabling detailed inspection for insurance claims, rental returns, and fleet management.
The Challenge
Gate cameras capture vehicles under varying lighting and angles. Damages can be subtle or severe, appearing at arbitrary orientations (e.g., diagonal scratches, bent panels, cracked lights). Standard axis‑aligned bounding boxes are inaccurate for such cases. The client required a fine‑grained classification with 11 damage categories and rotation‑aware localisation.
The Nazru Solution
Nazru’s platform integrated gate camera imagery with an oriented object detection deep learning model. Our algorithms automatically detect, localise (with oriented boxes), and classify each damage into one of the following classes:
Class | Description |
|---|---|
u bend | Deformation or bending of a panel (e.g., bumper, door edge) |
u crack | Linear fracture on rigid parts (e.g., bumper, light housing, windshield) |
u dent | Localised depression (small to large) on body panels |
u flat tyre | Tyre with no or very low air pressure |
u glass damage | Chips, cracks, or shattering on windows or windshield |
u light damage | Cracked, broken, or non‑functional headlight/taillight lens |
u missing part | Component completely absent (e.g., mirror cover, grille piece, emblem) |
u rim damage | Scratches, dents, or cracks specifically on wheel rims |
u rust | Corroded area on metal body panels |
u scratch | Thin, elongated abrasion on paint, glass, or trim |
u tyre damage | Tyre sidewall cuts, bulges, or tread damage (excluding flat) |
Each output includes: damage class, confidence, box parameters (center, width, height, angle), and optional associated vehicle part.
Key Results & Benefits
11 fine‑grained damage classes – covering body, glass, lights, tyres, and rims.
Oriented bounding boxes – precise fit for rotated damages (scratches, cracks, bends).
Supports automated inspection pipelines – rental returns, pre‑purchase checks, insurance claims, and fleet maintenance.
Integration with car part segmentation – map each damage to a specific part (door, bumper, wheel, etc.).
High accuracy with deep learning tailored to gate camera viewpoints.
Key Technologies Used
Oriented object detection (YOLOv8‑OBB, Rotated R‑CNN, or equivalent).
Multi‑class classification – 11 damage types trained on annotated gate images.
Angle regression – accurate rotation prediction for elongated damages.
Data augmentation – synthetic damages at various angles.
Output Formats (Rotating Rectangles)
Primary output: Oriented bounding boxes (GeoJSON, Shapefile, CSV, COCO‑OBB)
