30 credits – Machine Learning for Autonomous Driving at AI-Technologies Group

Scania genomgår nu en transformation från att vara en leverantör av lastbilar, bussar och motorer till en leverantör av kompletta och hållbara transportlösningar.

At the department of

Autonomous Transport Systems (ATS)

, we develop a full stack prototype system for sustainable autonomous driving. The work is done in close cooperation with Volkswagen group, leading technology suppliers and academic institutions. We take responsibility for creating a better and greener environment by finding answers to the questions on how we can achieve radically reduced emissions, manage the negative effects of ongoing urbanisation and save human lives in traffic.

A master thesis project in the AI Technologies group at ATS research, is a great opportunity to work on the forefront of autonomous vehicle development and machine learning technologies, that help to enable automated driving in complex driving environments. Many of our current employees have started their careers with a thesis project. We also have plenty of opportunities for both PhD and expat positions.

Background

The AI Technologies group is responsible for the development of Artificial Intelligence based functions for our autonomous driving pipeline. Most of our work currently circle around computer vision and deep learning, but we have plenty of ground to cover also outside of these domains.

Assignment

Below are two topics we are interested in exploring within the AI Technologies group. In both proposals you will as part of the thesis develop neural networks using our deep learning development infrastructure based on TensorFlow and Keras. Our pipeline is prepared for testing the final algorithms on the target system in our prototype self-driving vehicle, this is the goal for all our work including thesis projects. Depending on the successful applicant’s background, strengths and interests, we are flexible in the exact definition of the thesis.

Thesis 1: Object Separation using Semantic Segmentation and Bounding Boxes for Image Data

  • .
  • Segmentation networks provide object classification on pixel-level, but cannot separate overlapping objects. Also, object detection using bounding boxes can separate overlapping objects, however, lacks the pixel-level precision.
  • State-of-the-art instance segmentation networks that separate different object instances, such as the
  • Mask-RCNN

  • , has proven to be powerful but inefficient for real-time applications. By using real-time and efficient semantic segmentation networks, as well as object detection networks, we would like to have a method which combines the output from the two networks, separate object instances on pixel level, and increase the confidence level.
  • We are also in need of good
  • key performance indicators

  • (KPIs), that measures the performance of the proposed method. Applications in the autonomous driving domain are in need of robust performance metrics in general. Some standard metrics in segmentation networks, for example
  • accuracy

  • and
  • mIoU

  • , are sufficient within the academic work, but often come up short in complex applications.
  • Thesis 2: Real-Time Uncertainty Estimation for Semantic Segmentation Neural Networks.

  • In this thesis you will explore robust and efficient methods that can measure prediction uncertainty.
  • Real-time uncertainty estimation in machine learning applications play a vital role today, especially for a safety critical application like autonomous driving. When fusing data from different sources to prepare for autonomous decision making, each data source needs to have a probability, or a “certainty” measurement. State-of-the-art methods such as
  • Monte Carlo Dropout

  • (MC Dropout) are inefficient for real-time on-board applications. There are also studies that use the
  • softmax

  • and
  • entropy

  • for real-time estimation of the uncertainty, however they have drawbacks such as sometimes being certain for false-positive examples.
  • We are also in need of good
  • key performance indicators

  • (KPIs), that measures the performance of the proposed method. Applications in the autonomous driving domain are in need of robust performance metrics in general. Some standard metrics in segmentation networks, for example
  • accuracy

  • and
  • mIoU

  • , are sufficient within the academic work, but often come short in complex applications.
  • Education and skills

    Master student (Civilingenjör) in computer science, mathematics, physics or similar, preferably with specialization in computer vision, machine learning, artificial intelligence, data science, and robotics.

    Documented experience and skills in Python and C++, in addition to machine- and deep learning, is a merit.

    The work will be carried out at our offices in Södertälje. A thesis project is a great way to learn more about Scania and our many interesting career opportunities.

    Number of students: one student per topic (two in total)

    Start date: Autumn 2019 or Spring 2020

    Estimated time needed: 20 weeks, Full-time Language of work: Good knowledge in English is required

    Contact persons and supervisors

    Mikael Johansson, AI Team Leader, 08 553 533 45

    Addi Djikic, Supervisor, 08 553 722 68

    Ezeddin Al Hakim, Supervisor, 08 553 534 91

    Enclose CV, personal letter (also mentioning the preferred thesis) and grades.

    Skicka din ansökan till med rubrikraden Ny Teknik Jobb.

    Aktuellt inom