基于神经隐式表面表征的动态可扩展SLAM技术  

Dynamically scalable SLAM based on neural implicit surface characterization

作　　者：张嘉钏 戈湑 王录涛[1] ZHANG Jia-chuan;GE Xu;WANG Lu-tao(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《计算机工程与设计》2025年第4期1119-1126,共8页Computer Engineering and Design

基　　金：四川省科技计划重点研发基金项目(23ZDYF0756)。

摘　　要：针对同时定位与地图构建过程中,受遮挡、测量误差等因素影响导致构造地图易出现空洞、断层等问题,提出一种基于神经辐射场的三维场景重建方法,提升建图质量,实现无边界场景的重建和地图的动态扩展。基于密度场是由一个定义良好的表面产生的这一假设,通过约束表面的不透明度误差,使射线采样点更好地吻合辐射场分布;为实现压缩网格占用的空间的同时,不降低重建分辨率,引入线性插值的方法,获得指定坐标的特征向量集;引入光度约束、深度约束和符号距离值约束,提升复杂几何表面细节恢复能力。多个数据集与实际数据验证结果表明,系统在重建地图的精度、完整度和完整率上都有所提升,验证了方法的可行性。Aiming at the problems such as holes and faults in the constructed map caused by factors such as occlusion and mea-surement errors in the process of simultaneous localization and mapping,a 3D scene reconstruction method based on neural radiation field was proposed to improve the quality of map construction,and to realize the reconstruction of non-boundary scenes and the dynamic expansion of maps.Based on the assumption that the density field is generated by a well-defined surface,by limiting the opacity error of the surface,the sampling point of the ray was made better fitting the radiation field distribution.To achieve the compression of the space occupied by the mesh without lowering the reconstruction resolution,the linear interpolation method was introduced to obtain the set of eigenvectors with the specified coordinates.Luminosity constraint,depth constraint and symbolic distance value constraint were introduced to improve the detail recovery ability of complex geometric surfaces.The verification results of multiple data sets and actual data show that the system improves the accuracy,integrity and integrity rate of the reconstructed map,which verifies the feasibility of the method.

关 键 词：同时定位与地图构建 三维重建 神经隐式网络 体渲染 截断有符号距离函数 线性插值 体素网格 

分 类 号：TP242.62[自动化与计算机技术—检测技术与自动化装置]