World Model Online Evaluation

Abstract

To further advance the development of Advanced Driver Assistance Systems (ADAS), ensuring a robust learning-based perception system is essential. One approach to achieving this is the collection of diverse and informative data for training perception models. Active learning is based on the principle that exposing models to previously unseen or challenging data improves their performance. To facilitate this, this thesis evaluates several motion-based methods for detecting inconsistencies in the behavior of perception systems. These methods aim to identify scenarios in which the system performs poorly, enabling targeted data collection. The results show that, for a learning-based perception system, more complex methods generally yield better performance. In particular, Factor Graph Optimization with a Coordinated Turn motion model demonstrates the greatest potential, which may be attributed to the complex and dynamic behavior of objects in traffic environments. Furthermore, the findings suggest that future work exploring more advanced motion models, as well as combinations of multiple models, is a promising direction for improving perception system performance.