Unsupervised Monocular Road Segmentation for Autonomous Driving via Scene Geometry

TL;DR

This work tackles the costly reliance on labeled data for road-segmentation by introducing a fully unsupervised monocular framework that leverages scene geometry and temporal cues. It bootstraps learning with geometry-derived weak labels—identifying a horizon-based non-road region and a front-facing road quadrilateral—then refines predictions through temporal consistency via feature tracking and a mutual information loss, complemented by hard negative mining. The two-stage pipeline, implemented on a lightweight Lite R-ASPP architecture, achieves an IoU of $0.84$ on Cityscapes validation (and $0.82$ on test) without manual annotations, narrowing the gap to supervised methods while remaining computationally efficient. These results demonstrate the viability of combining geometric priors with temporal information for scalable unsupervised road segmentation and highlight its potential for robust deployment in diverse driving environments.

Abstract

This paper presents a fully unsupervised approach for binary road segmentation (road vs. non-road), eliminating the reliance on costly manually labeled datasets. The method leverages scene geometry and temporal cues to distinguish road from non-road regions. Weak labels are first generated from geometric priors, marking pixels above the horizon as non-road and a predefined quadrilateral in front of the vehicle as road. In a refinement stage, temporal consistency is enforced by tracking local feature points across frames and penalizing inconsistent label assignments using mutual information maximization. This enhances both precision and temporal stability. On the Cityscapes dataset, the model achieves an Intersection-over-Union (IoU) of 0.82, demonstrating high accuracy with a simple design. These findings demonstrate the potential of combining geometric constraints and temporal consistency for scalable unsupervised road segmentation in autonomous driving.

Figures (3)

• Figure 1: Overall framework of the proposed approach: initial training with partial mask and subsequent refinement with feature extraction and temporal consistency constraints. Feature extraction consists of corner detection and tracking across frames. The tracked corners are then filtered through hard negative mining before being fed into the IIC loss Ji2019IIC. Finally, the IIC loss is combined with the cross-entropy loss to form the final loss for training the segmentation network.
• Figure 2: A frame from the Cityscapes dataset with its partial mask and point tracking results. The green area represents the road, while the red area denotes non-road regions. Selected pairs of tracked points with inconsistent labels predicted by the network after the 10th epoch are highlighted.
• Figure 3: Examples of segmented Cityscapes images using the presented method