Depth estimation on wet roadways using RGB stereo

Computer vision systems are a key component for autonomous vehicles and driver assistance. To be
robust, such systems must be able to handle challenging conditions such as night driving and rain.
Previously Algolux proposed multimodal sensor fusion using RGB, gated and lidar for robust object
detection and dense depth estimation under low light and adverse weather. This project will address the
problem of wet roadways which produce elongated highlights that we call ‘specular streaks’. Specular
streaks can cause problems for stereo depth estimation because they give rise to disparity estimates that
are inconsistent with the roadway depth. The goal of this project is to segment these streaks and to infer
the roadway depth in their presence. The project will explore a combination of supervised, semisupervised
and self-supervised learning, and will combine learning-based approaches with physically
based image formation models process to enforce consistency between different prediction tasks.
Evaluation will be performed using manually annotated streak segmentation, lidar-based depth data, and
simulated data under different weather and illumination conditions in urban and highway scenarios.

Faculty Supervisor:

Michael Langer

Student:

Partner:

Algolux;Torc CND Robotics Inc.

Discipline:

Computer science

Sector:

Manufacturing; Professional, scientific and technical services

University:

McGill University

Program: