基于神经渲染的自动驾驶场景解耦表征与感知数据生成OA

High-quality perception data is crucial for designing and testing autonomous-driving algorithms.To address the limited scenario coverage of traditional data collection caused by high costs and the scarcity of long-tail data,the paper proposes a 4D scenarios modeling approach based on Neural Radiance Fields(NeRF).This method uses self-supervised learning to separate static and dynamic elements,and utilizes controllable appearance encoding to decouple the geometric and appearance features of the scenarios.Without retraining the model,diverse perception data can be generated simply by embedding appearance feature vectors.Tests on the Waymo dataset show that the perception data generated by this method achieves high realism,with a PSNR of 27.75 and an SSIM of 0.808,outperforming both D2NeRF and HyperNeRF.