Motivation

The perception of autonomous vehicles relies on a variety of sensors (cameras, LiDAR, radar) and AI algorithms for object detection, classification, and motion prediction. However, these systems are not perfect—uncertainties arise due to noise, poor visibility, limited training data, or model errors. For example, in 2016, a Tesla collided with a stationary truck because the system confused the white trailer tarp with the bright sky. Had the vehicle been able to assess its own uncertainty, the accident might have been preventable [1].
For autonomous vehicles to be not only safe but also certifiable, it is crucial to understand how uncertainties propagate within the perception chain (uncertainty propagation) and how they can be quantified in order to enable robust decision-making.

Tasks

The goal of this thesis is to conduct a structured literature review on approaches to uncertainty modeling and propagation in the perception of autonomous vehicles.

Specifically, the tasks includes:

• Systematic search and selection of relevant scientific publications
• Categorization of uncertainties (e.g., epistemic vs. aleatoric) in perception systems
• Analysis of existing methods for uncertainty quantification in
• Single sensors (e.g., probabilistic camera models, LiDAR error modeling)
  • Sensor fusion (e.g., Bayesian filtering, deep fusion with uncertainty estimation)
  • Object detection and tracking (e.g., Monte Carlo dropout)
  • Comparison of approaches for propagating these uncertainties throughout the entire perception pipeline
• Discussion of how these uncertainties can be integrated into safety arguments

The results will directly contribute to ongoing research on uncertainty modeling for safety arguments and certification processes.

Your profile

• Current studies in mechanical engineering, mechatronics, industrial engineering, or similar fields
• Basic knowledge of sensors, machine learning, or statistics is an advantage
• Interest in autonomous driving and probabilistic methods
• Analytical thinking and ability to present complex content in a structured manner

We offer

• Work on an exciting topic with high relevance
• Deep insights into the development and validation of autonomous driving functions
• Opportunity to make a meaningful contribution to current research
• Possibility of joint scientific publication
• Participation in a research project with renowned industry partners

Earliest starting date: Immediately

Interested? Please submit your application with a CV and grade overview to anian.scheibel∂kit.edu