Car Part Segmentation from Gate Camera Imagery (Case Study)
Location: (Gate camera)
Project Overview
Nazru’s AI platform was deployed to perform fine‑grained segmentation of vehicle parts from high‑resolution images captured by a gate camera. Each vehicle part was detected as a polygon and attributed with its part class and direction (e.g., left/right, front/rear). This enabled automated vehicle inspection, damage assessment, part‑level analytics, and security applications.
The Challenge
Gate cameras provide a controlled viewpoint as vehicles pass slowly or stop, but variations in lighting, reflections, occlusions (dirt, accessories), and diverse vehicle models (sedan, SUV, truck, hatchback) make accurate part segmentation difficult. Moreover, distinguishing between left and right parts (e.g., left door vs. right door, left headlight vs. right headlight) is essential for many use cases but is often overlooked. Traditional methods lacked the granularity and directional awareness required.
The Nazru Solution
Nazru’s platform integrated gate camera imagery with advanced instance segmentation deep learning models. Our algorithms automatically:
Detected the vehicle and isolated it from background.
Segmented each vehicle into its constituent parts with class‑ and direction‑aware labels.
Assigned a direction attribute (
left,right,front,rear, orcenter) to each part.Generated polygon masks for every part instance.
The following part classes (with direction variants) were used (example list – client to confirm):
Part Class | Direction Variants |
|---|---|
door | left_front, right_front, left_rear, right_rear |
window | left_front, right_front, left_rear, right_rear, front_windshield, rear_windshield |
wheel | left_front, right_front, left_rear, right_rear |
headlight | left, right |
taillight | left, right |
turn_signal | left, right (front/rear optional) |
mirror | left, right |
bumper | front, rear |
hood | center (front) |
trunk | center (rear) |
grille | center |
license_plate | center (front or rear) |
roof | center |
side_skirt | left, right (optional) |
fender | left_front, right_front, left_rear, right_rear (optional) |
Each polygon output includes attributes: part (base class), side (left/right/front/rear/center), confidence, area_px (and area_m² if calibrated).
Key Results & Benefits
Pixel‑precise car part polygons from gate camera images, with direction awareness.
Enables directional inspection – e.g., left‑side damage vs. right‑side damage, left headlight failure, right mirror missing.
Supports automated vehicle analytics – part presence/absence, alignment, wear, and aftermarket modifications.
Works across vehicle makes and models with consistent camera viewpoint.
Polygon outputs directly usable for geometric measurements (e.g., door gap width, wheel diameter).
Scalable deployment – model can be retrained for different gate angles (front, rear, side).
Key Technologies Used
Deep learning instance segmentation (e.g., Mask R‑CNN, YOLOv8‑seg, or ViT‑based models) trained on annotated gate camera images.
Direction‑aware classification – separate output heads or combined class labels (e.g.,
left_front_door).Polygon vectorisation – converting pixel masks to smooth, accurate polygons with sub‑pixel boundaries.
- Occlusion and reflection handling – robust data augmentation and model training.
