Applied AI / Sensor fusion

BEV road corridor segmentation

A sensor-fusion baseline that converts Argoverse 2 camera, LiDAR, ego pose, calibration, and HD map data into BEV inputs for main road corridor segmentation.

Status Public repository and technical report

Stack Python, PyTorch, Argoverse 2, OpenCV, NumPy, Matplotlib

Input data Camera images, LiDAR sweeps, calibration, ego pose, and HD map geometry

Output BEV datasets, segmentation baselines, prediction figures, and failure analysis

BEV road corridor segmentation examples showing input, target, raw prediction, and postprocessed prediction — Figure from the BEV road corridor segmentation repository.

Project notes

This page keeps the useful project details in one place: data preparation, target definition, baseline results, and limitations.

Predict the main road corridor in bird's-eye view.
The goal was a focused segmentation baseline using Argoverse 2 camera, LiDAR, ego pose, calibration, and HD map geometry.
A BEV data pipeline and segmentation baseline.
The pipeline aligns LiDAR and camera data into the ego frame, samples RGB from seven ring cameras, rasterizes map-derived corridor targets, trains PyTorch baselines, and writes figures for review.
Main corridor instead of all drivable area.
Early experiments with full drivable-area labels were too broad for the baseline. The final target uses ego-aligned lane boundary geometry to focus on the road corridor ahead.
0.8637 raw IoU; 0.8649 postprocessed IoU.
The final baseline used 320 samples, an 80/20 train-validation split, and 11 BEV channels combining LiDAR, RGB, distance, angle, and ego-position features.
Harder on angled roads and farther-forward regions.
The report identifies where the baseline works well and where sparse observations or road geometry make the result less reliable.
The repo includes code, figures, training histories, and analysis notes.
Argoverse 2 dataset files are not included. Data and derived figures are subject to the Argoverse license terms.