如何用Python、Numpy和PyVista移除点云红色边框内凹陷区域？

针对已移除凹陷边缘（红色区域）但无法识别平滑凹陷内部点的问题，以下是三种基于Python、Numpy、PyVista的可行方案：

方法一：投影平面+多边形内点判断（适合平面类凹陷） #

如果红色边缘构成闭合平面轮廓，可将点云投影到该平面，通过2D多边形内点检测筛选目标区域：

import pyvista as pv
import numpy as np
from sklearn.decomposition import PCA
from shapely.geometry import Point, Polygon

# 读取点云（复用现有代码）
point_cloud = pv.read('dataset/car_door_part_3.stl')
curvature = point_cloud.curvature()
threshold_curvature = 0.02
edge_indices = np.where(curvature > threshold_curvature)[0]
edge_points = point_cloud.points[edge_indices]

# 1. 对边缘点做PCA，确定投影平面
pca = PCA(n_components=2)
edge_proj = pca.fit_transform(edge_points)
proj_matrix = pca.components_.T
all_points_proj = point_cloud.points @ proj_matrix

# 2. 用边缘点投影构建闭合多边形
polygon = Polygon(edge_proj)

# 3. 判断每个点的投影是否在多边形内部
mask = np.array([Point(p).within(polygon) for p in all_points_proj])

# 4. 删除内部点，保留外部点
filtered_points = point_cloud.points[~mask]
filtered_cloud = pv.PolyData(filtered_points)

# 可视化结果
p = pv.Plotter()
p.add_mesh(filtered_cloud, color='white')
p.add_mesh(pv.PolyData(edge_points), color='red', point_size=5)
p.show()

方法二：区域生长法（适合连续邻域的凹陷） #

以已提取的边缘点为种子，基于邻域距离和法向量相似度向内生长，标记凹陷内部点：

import pyvista as pv
import numpy as np
from sklearn.neighbors import KDTree

# 读取点云及边缘点（复用现有代码）
point_cloud = pv.read('dataset/car_door_part_3.stl')
curvature = point_cloud.curvature()
threshold_curvature = 0.02
edge_indices = np.where(curvature > threshold_curvature)[0]
edge_points = point_cloud.points[edge_indices]

# 1. 计算点云法向量
point_cloud.compute_normals(inplace=True)
normals = point_cloud.point_normals

# 2. 构建KDTree用于邻域查询
kdtree = KDTree(point_cloud.points)

# 3. 初始化生长标记
visited = np.zeros(point_cloud.n_points, dtype=bool)
visited[edge_indices] = True
seed_indices = edge_indices.copy()

# 4. 区域生长参数
neighbor_k = 20  # 邻域点数
normal_angle_thresh = 0.5  # 法向量夹角阈值（弧度）
distance_thresh = 0.01  # 邻域距离阈值

while len(seed_indices) > 0:
    current_idx = seed_indices[0]
    seed_indices = seed_indices[1:]
    
    # 查询邻域点
    distances, indices = kdtree.query([point_cloud.points[current_idx]], k=neighbor_k)
    neighbor_indices = indices[0]
    
    for idx in neighbor_indices:
        if not visited[idx]:
            # 检查法向量相似度与距离
            normal_dot = np.dot(normals[current_idx], normals[idx])
            dist = distances[0][np.where(indices[0]==idx)[0][0]]
            if normal_dot > np.cos(normal_angle_thresh) and dist < distance_thresh:
                visited[idx] = True
                seed_indices = np.append(seed_indices, idx)

# 5. 删除生长标记的凹陷区域点
filtered_points = point_cloud.points[~visited]
filtered_cloud = pv.PolyData(filtered_points)

# 可视化
p = pv.Plotter()
p.add_mesh(filtered_cloud, color='white')
p.add_mesh(pv.PolyData(edge_points), color='red', point_size=5)
p.show()

方法三：PyVista内置包围点筛选（适合可构建闭合面的凹陷） #

先将边缘点生成闭合多边形网格，再用PyVista内置函数判断点是否在内部：

import pyvista as pv
import numpy as np

# 读取点云及边缘点（复用现有代码）
point_cloud = pv.read('dataset/car_door_part_3.stl')
curvature = point_cloud.curvature()
threshold_curvature = 0.02
edge_indices = np.where(curvature > threshold_curvature)[0]
edge_points = point_cloud.points[edge_indices]

# 1. 将边缘点生成闭合多边形网格
edge_poly = pv.PolyData(edge_points)
# 对边缘点进行三角化生成面
edge_poly = edge_poly.delaunay_2d()
# 拉伸成薄闭合体用于包围检测
enclosed_surface = edge_poly.extrude([0,0,0.01])

# 2. 判断每个点是否在闭合面内部
inside = point_cloud.select_enclosed_points(enclosed_surface, check_surface=False)
mask = inside['SelectedPoints'].astype(bool)

# 3. 删除内部点
filtered_points = point_cloud.points[~mask]
filtered_cloud = pv.PolyData(filtered_points)

# 可视化
p = pv.Plotter()
p.add_mesh(filtered_cloud, color='white')
p.add_mesh(edge_poly, color='red', opacity=0.5)
p.show()

注意事项 #

• 若边缘点无序，需先通过DBSCAN聚类后按角度/距离排序，确保多边形闭合且有序。
• 所有阈值参数（曲率、邻域数、距离、角度等）需根据点云实际尺度调整。

内容的提问来源于stack exchange，提问作者Shihan Chao