Scholarlink

<p>Building high-fidelity 3D maps from LiDAR scans in dynamic environments remains a challenging task due to ghosting artifacts caused by transient objects and geometric inconsistencies arising from sparse spatial observations. In this work, we propose a 4D implicit neural mapping framework that integrates localized temporal modeling with a structurally decoupled static–dynamic representation. We employed compactly supported Wendland radial basis functions as temporal bases to enforce strict temporal locality and eliminate long-range temporal leakage inherent in global basis functions such as the discrete cosine basis. Then, we propose a spatial context module that refines multi-resolution hash-encoded features through a residual multilayer perceptron to improve geometric coherence in sparsely observed regions. Finally, we proposed a dual-path decoder that separates time-invariant static geometry from time-variant dynamic deformation via independent prediction heads sharing a common backbone, enabling clean static map extraction via simple geometric thresholding on the decoupled representation. The proposed framework is trained end to end in an unsupervised manner. We evaluate our approach on CoFusion, Newer College, and KTH Dynamic Map Benchmark datasets and demonstrate improved reconstruction fidelity and dynamic segmentation accuracy over existing methods.</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

Localized basis functions for decoupled spatiotemporal neural distance fields

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Pragya Sankhla,
  • Deepanshu Singh Solanki,
  • Rajendra Nagar

摘要

Building high-fidelity 3D maps from LiDAR scans in dynamic environments remains a challenging task due to ghosting artifacts caused by transient objects and geometric inconsistencies arising from sparse spatial observations. In this work, we propose a 4D implicit neural mapping framework that integrates localized temporal modeling with a structurally decoupled static–dynamic representation. We employed compactly supported Wendland radial basis functions as temporal bases to enforce strict temporal locality and eliminate long-range temporal leakage inherent in global basis functions such as the discrete cosine basis. Then, we propose a spatial context module that refines multi-resolution hash-encoded features through a residual multilayer perceptron to improve geometric coherence in sparsely observed regions. Finally, we proposed a dual-path decoder that separates time-invariant static geometry from time-variant dynamic deformation via independent prediction heads sharing a common backbone, enabling clean static map extraction via simple geometric thresholding on the decoupled representation. The proposed framework is trained end to end in an unsupervised manner. We evaluate our approach on CoFusion, Newer College, and KTH Dynamic Map Benchmark datasets and demonstrate improved reconstruction fidelity and dynamic segmentation accuracy over existing methods.