Master Thesis - Uncertainty-aware road condition segmentation using DL

    Master Thesis - Uncertainty-aware road condition segmentation using DL

    Klimator is a Swedish software and intelligence company, that delivers state-of-the-art road weather data, empowering customers in different domains to capture and create real value. Klimator's mission is to enable data-driven change with cutting-edge technologies, with the vision of co-creating a better tomorrow. The company has an unparalleled legacy within road climatology and research creating its unique ability to unleash world-leading predictive and detective road weather data. With proprietary data platforms, Klimator converts inapplicable data to real and relevant intelligence that supports innovation within numerous industries. For example, automotive players utilize Klimators software and data to improve driver experience, enhance ADAS functionalities, and enable safe and scalable AD.

    Master Thesis project description

     

    Uncertainty-aware road condition segmentation using DL

     

    Background:
    Adverse road surface conditions (RSC) like water, snow, and ice increase the risk of accidents on the road. By providing autonomous vehicles with information about such conditions, they can adapt their speed, following distance, or choose alternative routes to reduce these risks. Additionally, unexpected or especially hazardous situations may require human intervention, and autonomous systems should be able to realize this and notify the driver.

     

    Description:
    This thesis aims to create an uncertainty-aware DL model that predicts the RSC in front of a moving vehicle while simultaneously estimating its predictive uncertainty, which should ideally be able to detect distribution shifts and out-of-distribution samples.
    To achieve this, a dataset is provided that contains two kinds of data: front-facing camera images and point-wise RSC estimates. The latter are measured underneath the vehicle, but they have been fused with the images through motion estimation, resulting in a large, sparsely annotated image segmentation dataset. The dataset composes many different vehicles in different parts of the world which lets us control the degree of distribution shift in holdout datasets.

     

    Methodology
    In this project, the student(s) will:

    1. Research existing methods for uncertainty-aware DL, with a focus on image segmentation​.
    2. Select or design one or more uncertainty-aware DL architectures suitable for the given task and train them on the provided dataset.
    3. Define suitable holdout datasets and evaluation methods and use them to assess the trained model's ability to estimate its own predictive uncertainty.

    Students
    The project is intended for 1–2 students with backgrounds in machine learning, and preferably computer vision. Experience with data wrangling, Python, and libraries like PyTorch, is a bonus.

    Supervisors
    Pontus Andersson and Gustaf Gulliksson, Klimator AB.

    Contact
    gustaf.gulliksson@klimator.se

     

     

    Note:

    We appreciate your interest in Klimator. Klimator is an equal employment opportunity employer, committed to maintaining a supportive and inclusive workplace for all employees. Klimator does not discriminate against, and prohibits harassment of, any applicant or employee based on race, color, sex, sexual orientation, gender identity, religion, national origin, age or disability.

    Apply now

    We’d love to hear from you. Please fill out this form.

    Passionate people push for progress

    We seek people who share our values and push us forward with energy and innovative thinking.