Automotive Radar Modeling for Virtual Simulation Based on Mixture Density Network

Road safety is the fundamental purpose of advanced driver assistance systems (ADAS), and automotive Radar often plays a significant role in reliable environment perception. Therefore, as the complexity of system integration continues to increase, the development quality and speed become increasingly crucial. Nowadays, sensor virtualization helps expedite the development process. Since Radar must perform several measurements in a short period to resolve ambiguity, the resulting radar signal is multi-dimensional and data-intensive. Therefore, analyzing these signals and generating models from them are not an easy task. To overcome the challenges, we present a Radar model based on Mixture Density Network (MDN) to generate production sensor errors that exhibit varied input correlations. Meanwhile, the errors modify a specific object list of Radar during the virtualized sensing process. The results show that MDN networks can better express the uncertainty of sensor detection errors and can be generalized to a generic sensor model. Finally, the MDN-based radar model is integrated into a multi-body simulation virtual platform with an open simulation interface to validate its performance.