High-fidelity automotive radar modeling method for few-shot new-scene adaptation

Automotive radar is a critical sensor in autonomous driving perception systems, and its corresponding radar model is an essential component for supporting autonomous driving testing. Compared with mechanism-based modeling methods that suffer from low computational efficiency, data-driven radar models offer high efficiency. However, they suffer from significant prediction errors when radar missed detections occur, due to abrupt jumps in radar outputs. Moreover, when applied to new scenes, model performance often degrades substantially due to the scarcity of labeled data. To address these challenges, this paper proposes a high-fidelity automotive radar modeling method for few-shot new-scene adaptation. The method adopts a two-stage framework. In the first stage, a high-fidelity radar model is constructed in the source scene by introducing a target-state prediction branch to identify missed-detection states and correct the model outputs accordingly, thereby reducing the prediction errors caused by missed-detection-induced discontinuities. In the second stage, the source CRM-ANN is retained as a prior model, and a lightweight revised radar model (RRM) is trained using a small amount of labeled new-scene data to compensate for scene-specific differences; the source CRM-ANN and RRM jointly constitute GRM-NS. Real-vehicle experiments are conducted to construct radar datasets for model training and validation. Experimental results show that, in the source scene, CRM-ANN reduces the MSEs of all five radar output parameters by at least 63.4% compared with RM-ANN. In the new scene, GRM-NS reduces the MSEs of all five radar output parameters by at least 41.9% compared with the source CRM-ANN while using only one-tenth of the labeled new-scene data. The proposed two-stage framework enables the construction of high-fidelity radar models with few-shot cross-scene adaptation capability and is valuable for autonomous driving simulation testing and related vehicle engineering applications.