Large‑Scale Road Boundary & Route Extraction from EV Camera (50,000 Images)

Project scale: 50,000 images
Camera type: Forward‑facing dashboard camera on an electric vehicle (EV)
Output format: Polyline (road boundaries and route paths)

Project Overview

Nazru’s AI platform processed 50,000 images captured by an electric vehicle’s dashboard camera to automatically extract road boundaries and route paths (e.g., lane edges, roadside curbs, and the primary drivable route). All features were delivered as polylines for use in high‑definition mapping, autonomous driving, and route analytics.

The Challenge

With 50,000 images covering diverse road types (highways, rural roads, urban streets, curved/mountain roads) and varying conditions (day/night, rain, shadow, occlusion), manual annotation of road boundaries and routes is impossible at this scale. The client required a fully automated pipeline that:

• Extracts left and right road boundaries (curbs, edges, or lane lines).

• Delineates the drivable route path (e.g., centerline of the ego lane or the entire road).

• Outputs smooth, accurate polylines suitable for GIS and navigation.

• Handles occlusions (vehicles, trees) and challenging lighting.

The Nazru Solution

Nazru’s platform combined a scalable deep learning pipeline with efficient post‑processing. For each image, our algorithms:

• Detected road boundaries using semantic segmentation of lanes, curbs, and road edges.

• Generated a drivable route polyline by either:

  • Tracing the centre of the ego lane, or

  • Creating a smooth path between the left and right boundaries.

• Output polylines as ordered point sequences in image coordinates (later georeferenced with vehicle GPS/IMU).

• Processed all 50,000 images in a fully automated batch, with optional quality control (e.g., uncertain flags for low‑confidence segments).