Visual Anomaly Detection in Autonomous Driving
Description
Detecting objects in front of a self-driving vehicle is critical for safety but poses significant challenges. The range of potential objects on the road includes rare and unknown entities—such as wild animals, debris, or litter—which are underrepresented in existing datasets. The wide variety of these unknowns in terms of appearance, size, and location makes detection even more difficult.
Additionally, deep learning models are typically designed under the closed-set assumption, where they are trained on a fixed set of object classes. Although effective on curated datasets, these models often struggle in real-world open-set scenarios, where they frequently misclassify unknown objects with high confidence as belonging to known classes.
The aim of this thesis is to develop a method that expands the semantic segmentation capabilities of a vehicle’s image processing system to detect unknown objects, referred to as anomalies, in front of the vehicle.
Prerequiste
Proficiency in Python and relevant machine learning libraries, such as PyTorch, is essential. A solid understanding of machine learning and image processing is also required.