PIPER-AD

Prof. Dr.-Ing. Antje Muntzinger, Professor for Computer Vision at HFT Stuttgart, is investigating the development of a multimodal AI system for the robust detection and tracking of vehicles on highways. She combines AI with prior physical knowledge to compensate for sensor errors.

Modern AI models for detecting the surroundings of autonomous vehicles deliver impressive results. However, they are often sensitive to data noise, occluded objects and difficult visibility or weather conditions. Implausible results can occur, especially with single-image predictions, which often have to be corrected using downstream processes. However, highly automated driving systems require a robust and reliable perception of other road users through the combination of different sensor modalities.

The "PIPER-AD" project aims to develop an AI-supported perception system for highly automated driving systems. It is pursuing a hybrid approach by combining multimodal deep learning models with prior physical knowledge. Physical framework conditions, such as realistic vehicle dynamics and movement along the road, are integrated into the training process in order to increase the robustness and generalizability of the system. In the long term, the framework is to be expanded to include uncertainty quantification in order to increase safety even under difficult conditions and the acceptance of autonomous vehicles.

The goal of this research project is to develop an AI-based perception system for autonomous systems that uses novel physical regularizations to generate physically plausible results. The physically informed neural networks developed are intended to be transferable to other application areas in the future and to demonstrate the fundamental possibility of integrating prior physical knowledge into environmental perception.