[Paper Review] SafeDrive: A Robust Lane Tracking System for Autonomous and Assisted Driving Under Limited Visibility

Junaed Sattar, Jiawei Mo|arXiv (Cornell University)|Jan 29, 2017

Autonomous Vehicle Technology and Safety10 references3 citations

TL;DR

SafeDrive proposes a low-cost, vision-based lane tracking system that enhances robustness under poor visibility by leveraging vehicle location to retrieve past road-view imagery from a database, then aligns this imagery with real-time camera feeds using feature-based and pixel-based homography estimation to detect lane markers even when they are invisible in the current scene. The system reduces SSD error by 25% compared to baseline methods, achieving accurate lane projection in degraded conditions.

ABSTRACT

We present an approach towards robust lane tracking for assisted and autonomous driving, particularly under poor visibility. Autonomous detection of lane markers improves road safety, and purely visual tracking is desirable for widespread vehicle compatibility and reducing sensor intrusion, cost, and energy consumption. However, visual approaches are often ineffective because of a number of factors, including but not limited to occlusion, poor weather conditions, and paint wear-off. Our method, named SafeDrive, attempts to improve visual lane detection approaches in drastically degraded visual conditions without relying on additional active sensors. In scenarios where visual lane detection algorithms are unable to detect lane markers, the proposed approach uses location information of the vehicle to locate and access alternate imagery of the road and attempts detection on this secondary image. Subsequently, by using a combination of feature-based and pixel-based alignment, an estimated location of the lane marker is found in the current scene. We demonstrate the effectiveness of our system on actual driving data from locations in the United States with Google Street View as the source of alternate imagery.

Motivation & Objective

• To improve lane tracking reliability in autonomous and assisted driving under poor visibility conditions such as snow, low light, or occlusion.
• To eliminate reliance on expensive or intrusive sensors like LiDAR or radar by using only visual data and GPS.
• To develop a low-cost, smartphone-compatible solution for lane departure warnings in consumer vehicles.
• To enable lane detection in scenarios where visual cues are completely absent or severely degraded.
• To demonstrate the feasibility of using historical road-view imagery as a fallback for real-time visual lane detection.

Proposed method

• The system uses vehicle GPS coordinates to retrieve a geotagged, past image of the same road location from a database such as Google Street View.
• Feature-based matching (e.g., SIFT or similar) is applied to identify corresponding points between the current camera image and the retrieved database image.
• A homography matrix is estimated from the matched feature points to align the database image with the current scene.
• Pixel-based image alignment using the Coherent Point Drift (CPD) or similar algorithm refines the homography to reduce projection errors.
• A common pixel region is extracted based on matched features, and a final SSD (Sum of Squared Differences) metric is computed to validate alignment quality.
• Lane markers are detected in the aligned database image and projected back onto the current camera image using the refined homography for real-time lane visualization.

Experimental results

Research questions

• RQ1 Can lane detection be reliably performed when visual cues such as lane markers are completely invisible due to weather or lighting conditions?
• RQ2 How effectively can historical road-view imagery be used to recover lane information in real-time driving scenarios?
• RQ3 What is the performance gain of combining feature-based and pixel-based alignment in homography estimation for lane marker recovery?
• RQ4 How does the system handle geometric distortions and viewpoint differences between the current and past images?
• RQ5 Can a low-cost, GPS-assisted visual system achieve sufficient accuracy for real-time lane departure warnings in consumer vehicles?

Key findings

• The system successfully detects and projects lane markers onto the current image even when they are completely invisible, as demonstrated in snow-covered and low-light conditions.
• The SSD (Sum of Squared Differences) error between the current image and the projected database image was reduced from 42,603.4 to 31,576.6 after pixel-based refinement, indicating improved alignment accuracy.
• The final homography matrix after pixel-based refinement (matrix 3) showed significant improvement over the initial feature-based matrix (matrix 2), with reduced projection errors and better alignment of structural features like building edges.
• The average end-to-end processing time for lane detection and projection is approximately 8 seconds on a standard PC, making the system feasible for real-time deployment.
• The method effectively handles large viewpoint differences and scene variations between the current and database images, as validated in multiple test cases with different lighting and weather conditions.
• The system demonstrates that combining GPS-based image retrieval with multi-stage image alignment enables robust lane tracking in extreme visibility degradation, offering a viable low-cost alternative to sensor-heavy systems.