Comparison of Different Adaptive Light Distributions for Automated Driving

The contribution at hand compares nine innovative adaptive luminous intensity distributions (LIDs) for automated driving, as human-focused LIDs are mostly inefficient for computer vision. The adaptive LIDs are categorized into object-based and material-based LIDs. The advantages and disadvantages, as well as the requirements for the functionality of each LID, are discussed. All adaptive LIDs plus two reference LIDs are evaluated for camera object detection with neural networks in a worst-case scenario. Additionally, the robustness against object position or localization errors is tested. The results show that material-based LIDs outperform all others, e.g., by an improvement in detection confidence of 24% compared to one variant of object-based lighting. When considering inaccuracies, the object-based LIDs yield poor results. In contrast, those that only require a static 3D environment model show an improvement of 186% in terms of confidence and 178% in terms of Intersection over Union (IoU) compared to the best-scoring object-based LID.