Autonomous driving has made remarkable progress over the past decade, particularly in road scene analysis, thanks to the rise of deep learning approaches. However, the deployment of fully autonomous vehicles remains limited by the lack of robustness of perception systems, which are vulnerable to adverse weather conditions or sensor failures. These systems do not include explicit mechanisms to represent and reason about uncertainty. Achieving reliable perception therefore requires new frameworks capable of both integrating information from heterogeneous sensors and quantifying the uncertainty associated with their predictions. In this thesis, we address these challenges by proposing adaptive multimodal fusion strategies based on Dempster–Shafer Theory, which provides tools to represent epistemic uncertainty and manage conflicts between different sources of information. The first part of the manuscript reviews the main sensors used for road scene analysis, as well as deep learning architectures for perception tasks and the datasets used to evaluate them. The fundamental principles of Dempster–Shafer Theory are then introduced, including information representation, information fusion, and decision-making. Existing evidential neural networks for computer vision are also presented. To address the challenge of integrating Dempster–Shafer Theory into neural networks, algorithmic optimizations are developed, making these networks applicable to large-scale datasets. The advantages of evidential networks over probabilistic approaches are highlighted in scenarios where epistemic uncertainty plays a major role, such as out-of-distribution data detection. This work also focuses on multimodal semantic segmentation. The proposed architecture leverages Dempster–Shafer Theory to adaptively manage conflicts between sensors, ensuring reliable semantic segmentation in case of sensor failures. Finally, this architecture is extended by replacing "mass functions" with "belief intervals" to achieve a finer-grained representation of uncertainty. New fusion methods adapted to "belief intervals" are also explored. Experiments on both real-world and synthetic datasets demonstrate that this method further improves performance under sensor failures and adverse weather conditions.
Lucas Deregnaucourt (Tue,) studied this question.