Détection d’objets inconnus en monde ouvert pour les systèmes de conduite autonome

Autonomous vehicle (AV) systems primarily perform two core tasks: perception and control. The perception system constructs a representation of the real world, commonly referred to as the World Model, on which the control system relies upon to make driving decisions. The effectiveness of the control system is therefore closely tied to the accuracy and completeness of the information provided by the World Model. For example, failing to detect critical objects on the road may result in dangerous outcomes, while excessive or erroneous detections can negatively affect the user experience. Discrepancies between the real environment and the AV’s internal representation are known as corner cases. Among these, unknown objects, objects belonging to classes not encountered during training, represent a particularly challenging type of corner case. Contemporary World Models are predominantly based on deep learning, utilizing large-scale labeled datasets to learn representations of known object classes. Object detection, a central component of the perception pipeline, aims to provide both localization and semantic information about objects in the scene. However, state-of-the-art detectors exhibit poor performance on corner cases, as demonstrated in real-world benchmarks such as CODA. This limitation stems from their closed-set assumption: they are trained on a fixed set of classes and struggle to generalize to novel, unseen instances. To address this, open-set recognition and open-world object detection (OWOD), have been proposed. These paradigms introduce mechanisms for detecting unknown objects while enabling incremental learning of new classes. Nevertheless, applying OWOD to AV contexts presents unique challenges that remain insufficiently addressed. Most existing approaches are evaluated on general-purpose datasets such as COCO, which do not adequately reflect the complexity and safety-critical nature of road scenes. Current methods often suffer from high false positive rates or degrade performance on known classes when optimizing for unknown detection. This PhD thesis aims to bridge these gaps by advancing object detection systems for autonomous driving scenarios, with a specific focus on robust detection of unknown objects. The goal is to enhance the AV’s ability to handle corner cases in real-world environments, therebycontributing to safer and more reliable autonomous driving systems.