基于模板替换的室内场景建模方法研究

Indoor scene modeling method based on template replacement

当前,室内场景建模相关研究已经取得很多进展,特别是基于多视角融合的建模框架与基于单视角的建模框架的提出,增强了机器人的环境感知能力。但仍然存在以下不足:①基于多视角融合的建模方式预处理时间长,建模完成后需线下优化过程,不能满足特定条件下的建模需求;②基于单视角的建模算法输出一般为体素,建模质量较低,信息缺失严重,对于场景细节无法精确刻画,难以满足机器人交互的要求。特提出一种基于模板替换的室内场景建模方法研究。首先,预处理由设备采集到的三维点云场景,分割出存在点云缺失的单个对象,并利用虚拟扫描技术采样对象表面点并计算法向量与曲率。采用八叉树网格结构,将点云的法向量与曲率信息分别存入网格中,再利用卷积神经网络(CNN)提取高维特征向量,将其与数据库中三维对象特征进行欧氏距离比较,得到检索序列。从序列中挑选出最相似的对象,利用迭代就近点(ICP)配准方法,与扫描场景进行配准,完成场景优化。对提出的网络模型在2个基准数据集上进行测试并表现出良好的性能。

Nowadays,much progress has been made in the research of indoor scene modeling,especially the modeling frameworks based on multiple perspectives and single perspective,which has enhanced the robot’s environment perception.However,the following shortcomings still exist:①The modeling method based on multiple perspectives requires a long pre-processing time,and the offline optimization process is required after the modeling is completed,which cannot meet the modeling requirements under specific conditions.②The modeling algorithm based on single perspective is mainly output with voxels,so the modeling quality is low,and the information is missing seriously.The details of the scene cannot be accurately characterised,and it is difficult to meet the requirements of robot interaction.In view of the above deficiencies,this paper puts forward a method of indoor scene modeling based on template replacement.First,the three-dimensional point cloud scene is preprocessed to segment a single object with missing point cloud,and then the virtual scanning technology is used to sample the surface points of the object and calculate the corresponding normal vector and curvature.Next,the octree mesh is used to store the normal vector and the curvature information respectively.Furthermore,the high-dimensional feature vectors are extracted by the convolutional neural network(CNN),and the Euclidean distance is compared with the features of three-dimensional object in the database,so as to obtain the retrieval sequence.Finally,the most similar objects are selected from the sequence,and the iterative closest point(ICP)registration method is used to register with the scanning scene to complete the scene optimization.In this paper,the proposed network model is tested on two benchmark data sets and shows good performance.