Performance evaluation of a state-of-the-art automotive radar and corresponding modeling approaches based on a large labeled dataset

Abstract

Radar is a key sensor to achieve a reliable environment perception for advanced driver assistance system and automated driving (ADAS/AD) functions. Reducing the development efforts for ADAS functions and eventually enabling AD functions demands the extension of conventional physical test drives with simulations in virtual test environments. In such a virtual test environment, the physical radar unit is replaced by a virtual radar model. Driving datasets, such as the nuScenes dataset, containing large amounts of annotated sensor measurements, help understand sensor capabilities and play an important role in sensor modeling. This article includes a thorough analysis of the radar data available in the nuScenes dataset. Radar properties, such as detection thresholds, and detection probabilities depending on object, environment, and radar parameters, as well as object properties, such as reflection behavior depending on object type, are investigated quantitatively. The overall detection probability of the considered radar (Continental ARS-408-21) was found to be 27.81%. Four radar models on object level with different complexity levels and different parametrisation requirements are presented: a simple RCS-based radar model with an accuracy of 51%, a linear SVC model with an accuracy of 70%, a Random Forest model with an accuracy of 83%, and a Gradient Boost model with an accuracy of 86%. The feature importance analysis of the machine learning algorithms revealed that object class, object size, and object visibility are the most important parameters for the presented radar models. In contrast, daytime and weather conditions seem to have only minor influence on the modeling results.