Real-Time Fully Unsupervised Domain Adaptation for Lane Detection in Autonomous Driving

TL;DR

This paper introduces a lightweight, fully unsupervised, real-time lane detection adaptation method for autonomous driving that adjusts only batch-normalization parameters, achieving high accuracy and real-time performance on embedded devices.

Contribution

The paper presents a novel real-time unsupervised adaptation technique that modifies only batch-normalization parameters, enabling on-device lane detection adaptation without labeled data.

Findings

Achieves 92.19% average accuracy in lane detection.

Operates at 30 FPS on Nvidia Jetson Orin.

Matches semi-supervised methods' accuracy without requiring labels.

Abstract

While deep neural networks are being utilized heavily for autonomous driving, they need to be adapted to new unseen environmental conditions for which they were not trained. We focus on a safety critical application of lane detection, and propose a lightweight, fully unsupervised, real-time adaptation approach that only adapts the batch-normalization parameters of the model. We demonstrate that our technique can perform inference, followed by on-device adaptation, under a tight constraint of 30 FPS on Nvidia Jetson Orin. It shows similar accuracy (avg. of 92.19%) as a state-of-the-art semi-supervised adaptation algorithm but which does not support real-time adaptation.