Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Master Thesis: Model Transfer / Domain Adaptation for Driver State Detection
Driver state detection systems will become mandatory in many countries in this decade. In this master thesis, you will work on the data differences and the model transferability between different domains (e.g., simulator and real vehicle) by using machine learning techniques.
Keywords: driver state detection; data analytics; explainable machine learning; transfer learning; domain adaptation
**Motivation**
_“New technology mandate in infrastructure bill could significantly cut drunken driving deaths”_ (The Washington Post, 09.11.2021)
In the US and Europe, driver state detection systems are becoming required by law soon. Our chair has already vast experience in research on driver state detection systems. To reach our goals, we closely collaborate with one of the globally leading automotive suppliers, Bosch GmbH, and the University of St. Gallen. To give one example, we conducted a field study on test track and were able to detect whether drivers with diabetes were in a critical health state of hypoglycemia state (too low blood sugar) or not. Check out the following video (in German): https://www.srf.ch/play/tv/redirect/detail/e8217a30-f359-416d-949a-051ee6a240e6
For the detection of hypoglycemia, we have also conducted simulator studies in the lab. Additionally, we will soon collect data in real traffic without any intervention. An overview of the three distinct levels of datasets can be found in the attached image.
**Motivation**
_“New technology mandate in infrastructure bill could significantly cut drunken driving deaths”_ (The Washington Post, 09.11.2021)
In the US and Europe, driver state detection systems are becoming required by law soon. Our chair has already vast experience in research on driver state detection systems. To reach our goals, we closely collaborate with one of the globally leading automotive suppliers, Bosch GmbH, and the University of St. Gallen. To give one example, we conducted a field study on test track and were able to detect whether drivers with diabetes were in a critical health state of hypoglycemia state (too low blood sugar) or not. Check out the following video (in German): https://www.srf.ch/play/tv/redirect/detail/e8217a30-f359-416d-949a-051ee6a240e6
For the detection of hypoglycemia, we have also conducted simulator studies in the lab. Additionally, we will soon collect data in real traffic without any intervention. An overview of the three distinct levels of datasets can be found in the attached image.
Studying drivers in real vehicles is much more effort than using a simulator – and doing studies in real traffic is too! Hence it would be helpful, if a driver state detection model could be trained on simulator data and then transferred into the real vehicle (while having no or limited labeled training data).
One option to solve such a problem could be domain adaptation: The goal is to deal with models that are trained on a source distribution and should then be used on a different target distribution (violet arrows in the figure). The tasks are the same in both domains (e.g., hypoglycemia detection in our case). Hence, it is a subclass of transfer learning.
**Working Packages**
The goal of this thesis is to approach the problem of generalizing from driver state prediction models from simulator datasets to ML models which also work in real vehicles. In doing so, we are not only aiming for a solid detection performance but also for high interpretability. Hence, the following steps should be done:
1. Analyze the data, find similarities and differences of the datasets and their features
2. Research on state-of-the-art domain adaptation techniques and decide what is best to solve our problem (trade-off interpretability vs. performance)
3. Implement your solution and asses the detection performance on our datasets
During the thesis, you will work closely with us and will receive dedicated supervision. We are highly interested to invest time in your thesis as the topic above is one on the top of our research agenda.
- _What you offer:_ master student, highly motivated, technical background, python skills, some machine learning background
- _What we offer:_ interdisciplinary expert team, unique data sets, very good working environment, publication option
Do not hesitate to apply! Please contact us with your CV and your up-to-date transcripts of records (bachelor and master). If available, an overview of completed projects can be an advantage.
Studying drivers in real vehicles is much more effort than using a simulator – and doing studies in real traffic is too! Hence it would be helpful, if a driver state detection model could be trained on simulator data and then transferred into the real vehicle (while having no or limited labeled training data).
One option to solve such a problem could be domain adaptation: The goal is to deal with models that are trained on a source distribution and should then be used on a different target distribution (violet arrows in the figure). The tasks are the same in both domains (e.g., hypoglycemia detection in our case). Hence, it is a subclass of transfer learning.
**Working Packages**
The goal of this thesis is to approach the problem of generalizing from driver state prediction models from simulator datasets to ML models which also work in real vehicles. In doing so, we are not only aiming for a solid detection performance but also for high interpretability. Hence, the following steps should be done:
1. Analyze the data, find similarities and differences of the datasets and their features 2. Research on state-of-the-art domain adaptation techniques and decide what is best to solve our problem (trade-off interpretability vs. performance) 3. Implement your solution and asses the detection performance on our datasets
During the thesis, you will work closely with us and will receive dedicated supervision. We are highly interested to invest time in your thesis as the topic above is one on the top of our research agenda.
- _What you offer:_ master student, highly motivated, technical background, python skills, some machine learning background - _What we offer:_ interdisciplinary expert team, unique data sets, very good working environment, publication option
Do not hesitate to apply! Please contact us with your CV and your up-to-date transcripts of records (bachelor and master). If available, an overview of completed projects can be an advantage.