Scholarlink

<p>This study addresses the limitations of Neural Radiance Fields (NeRF) and its derivatives in dynamic driving scenes by proposing an optimization method based on MARS to improve 3D reconstruction efficiency and stability in complex driving environments. Model parameters are initialized from pre-trained background, global, and object models. To optimize feature extraction and reduce overfitting, mean squared error and regularization terms are introduced. The improved optimizer employs an efficient gradient descent strategy and decouples weight decay from gradient updates. Through iterative backpropagation, model parameters are updated to optimize the loss function, enhancing the model's ability to recognize both static and dynamic objects. The method is evaluated on the KITTI and VKITTI datasets with different splits (Novel View Synthesis (NVS)-75%, NVS-50%, and NVS-25%) and compared with NSG, SUDS, and MARS. The results show a 7.3% improvement in PSNR and a 28.9% reduction in LPIPS, with relatively better performance under low data ratio conditions, enhancing reconstruction quality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The 3D reconstruction method for driving scenes based on improved neural radiance fields

  • 期刊论文

  • 发表日期：

    2025年6月17日

查看全文
  • Weizhi Zhu,
  • Lianfang Tian,
  • Qiliang Du,
  • Juanhong Xie

摘要

This study addresses the limitations of Neural Radiance Fields (NeRF) and its derivatives in dynamic driving scenes by proposing an optimization method based on MARS to improve 3D reconstruction efficiency and stability in complex driving environments. Model parameters are initialized from pre-trained background, global, and object models. To optimize feature extraction and reduce overfitting, mean squared error and regularization terms are introduced. The improved optimizer employs an efficient gradient descent strategy and decouples weight decay from gradient updates. Through iterative backpropagation, model parameters are updated to optimize the loss function, enhancing the model's ability to recognize both static and dynamic objects. The method is evaluated on the KITTI and VKITTI datasets with different splits (Novel View Synthesis (NVS)-75%, NVS-50%, and NVS-25%) and compared with NSG, SUDS, and MARS. The results show a 7.3% improvement in PSNR and a 28.9% reduction in LPIPS, with relatively better performance under low data ratio conditions, enhancing reconstruction quality.