Lpd Net Large Scale Place Pointcloud Cls&seg

Posted by Packy on February 10, 2020

2019_ICCV_LPD-Net: 3D Point Cloud Learning for Large-Scale Place Recognition and Environment Analysis

Timeframe  Figure l. Place recognition in large-scale environments. We use  global feature descriptors generated from the raw 3D point cloud  data to accomplish place recognition tasks. The lower side shows  two examples with different weather and light conditions.

Point cloud based place recognition is still an open issue due to the difficulty in extracting local features from the raw 3D point cloud and generating the global descriptor, and it’s even harder in the large-scale dynamic environments.

![map Transform Local Feature Extraction Point Cloud Transform and Adaptive Local Feature Extraction I Feattre Neighb« Dynamic (KNN-m1p) Feature S Spatial Neighb« I Rela lions (KNN-m1p) Cutesian Space (64.12s.1024) mlp I loutput Global Global Feature Feature A I Graph-based Neighborhood Aggregation I Figure 2. LPD-Net Architecture. The network takes the raw point cloud data as input, applies Adaptive Local Feature Extraction to obtain the point cloud distribution and the enhanced local features, which are aggregated both in the Feature Space and the Cartesian Space through the graph neural network. The resulted feature vectors are then utilized by NetVIÄD l to generate a global descriptor.

问题:大规模的点云场景如何进行识别。

难点:如何对原始点云进行特征提取(点级别的),并且聚合为全局的描述符。

同时如何处理大规模的动态场景也是一个问题。

方法: 对于点云的特征提取, 利用了点云的局部几何特征和经过T-net变换后的点云坐标,(tnet来解决旋转不变性,局部几何特征为了充分考虑点云的统计特性);然后经过mlp后进行图卷积, 分别考虑了特征空间的knn图,和空域中的knn图。 应用了 lazy quadruplet loss

关于几何特征如何选择:{

适用于描述大型场景的局部特征可以分为四类:基于特征值的3D特征(F3D),将3D点投影到水平面上而产生的特征(F2D),基于法向矢量的特征(FV) 以及基于Z轴统计信息(FZ)的功能。 现有研究证实,F3D,FV和FZ可有效解决大规模3D场景分析问题[24],F2D和FZ可有效解决自动驾驶任务中的大规模定位问题[3,4]。 考虑到特征的冗余性和可分辨性,我们选择以下十个局部特征来描述每个点i周围的局部分布和结构信息:

}

3维特征: 曲率变化, Omni-variance全方差?,线性度Linearity,特征值熵,局部点密度

2维特征:2D scattering2维线性度2D linearity

FV特征:法线向量Vi的垂直分量

FZ特征: 最大高度差 高度方差

自适应的选择k:

我们通过考虑每个点i周围的局部3D结构来介绍局部分布特征。计算k个最近的邻近点,并将各个局部3D位置协方差矩阵视为局部结构张量。在不失一般性的前提下,我们假设λi 1≥λi 2≥λi 3≥0代表对称正定协方差矩阵的特征值。根据[24],以下测量可用于从香农信息熵理论的角度描述局部结构的不可预测性,Ei = -Li lnLi-Pi ln Pi-Si ln Si,(1)其中,Li = λi1-λi 2λi 1,Pi =λi 2-λi 3λi 1和Si =λi 3λi 1分别表示每个点局部邻域的线性,平面性和散射特征。这些特征描述了每个点周围的1D,2D和3D局部结构[24]。由于点云中的点分布通常是均匀的,因此我们通过最小化不同k值之间的Ei来自适应地选择每个点i的邻域,并将最佳邻居大小确定为k i opt = arg min k Ei(k)。

同时论文还讨论了怎么设计Feature Transformation提出了三种可能方法:

Feature Neighbor I  Relations I  F cat ure  (a) FN-Original structure (O)  Trans form  Neighbor Relations I  Feature  (b) FN-Series structure (S)  Feature  Neighbor Relations  Feature  (c) FN-Parallel structure (P)  Figure 3. Different network structures for feature transformation and relation extraction.

通过实验发现方式c取得的效果最好。 fn并行。

Similar local 3D structures which locate in different parts of the point cloud usually have similar local features. Their spatial distribution relationships are also of great importance in place description and recognition tasks.

即局部的几何特征和原始的空间关系都很重要。

所以在特征空间做动态图,原始的坐标空间固定的KNN图。

Feature Neighbor  Relations  Dynamic Graph  (KNN- )  Spatial Neighbor  Rela ions  Neighbor Graph  KNN-m1  Feature Space  Concatenate 1  Cartesian Space I  Feature Neighbor  Relations  Dynamic  Spatial Neighbor  Neighbor Graph  (KNN )  Feature Space I  Maxpooling  Cartesian Space I  I Feature Neighbor  Relations  Dynamic Graph  I (KNN-mlp)  Feature Space  Spatial Neighbor  Relations  Neighbor Graph  (KN N -mlp)  Cartesian Space  Graph-based Neighborhood Aggregation I  Graph-based Neighborhood Aggregation I  (a) Prarllel-Concatenation structure (PC) (b) Parallel-Maxpooling structure (PM)  I Graph-based Neighborhood Aggregation I  (c) Series-FC structure (SF)  Figure 6. Different network structures for feature aggregation.

提出了三种可能的方式,

创新点: 由于点云的不均匀的特性,所以要自适应的局部点特征提取,利用了香农信息论的一些东西,来最优化k,实现自适应选取邻域的范围。

这里提取的局部特征主要有:the 1D, 2D and 3D local structures around each point 等。

不足: 标题中的大规模,只是体现在原始点云的尺度上,而不是点云数量规模大。