Unknown-Aware Hierarchical Object Detection in the Context of Automated Driving

Recent research in data-driven perception models has yielded promising performance on the previously trained classes. However, in real-world driving scenarios, the robustness of automated driving functions also depends on their ability to react safely to the occurrence of object classes that were unknown at training time. To address this challenge, we propose an unknown-aware object detection approach that incorporates a-priori knowledge in the form of a class taxonomy integrated into the network. Our proposed method utilizes two flexible modules responsible for object localization and hierarchical classification to ensure that the network can detect both previously known and unknown classes, without using additional training data that includes unknown classes. Moreover, we discuss the impact of the selected class hierarchy and training methodology on the detection performance.