【原】Open3d学习计划—高级篇 6（体素化）

点云PCL 2021-03-09

展开全文

Open3D是一个开源库，支持快速开发和处理3D数据。Open3D在c++和Python中公开了一组精心选择的数据结构和算法。后端是高度优化的，并且是为并行化而设置的。

本系列学习计划有Blue同学作为发起人，主要以Open3D官方网站的教程为主进行翻译与实践的学习计划。点云PCL公众号作为免费的3D视觉，点云交流社区，期待有使用Open3D或者感兴趣的小伙伴能够加入我们的翻译计划，贡献免费交流社区，为使用Open3D提供中文的使用教程。

点云和三角网格是一种十分灵活的，但是不规则的几何类型。体素网格是通过规则的3D网格来表示的另一种3D几何类型，并且它可以看作是2D像素在3D上的对照物。Open3d中的VoxelGrid几何类型能够被用来处理体素网格数据。

从三角网格中生成

Open3d提供了create_from_triangle_mesh函数能够从三角网格中生成体素网格。它返回一个体素网格，其中所有与三角形相交的网格被设置为1，其余的设置为0。其中voxel_zie参数是用来设置网格分辨率。

print('input')mesh = o3dtut.get_bunny_mesh()# fit to unit cubemesh.scale(1 / np.max(mesh.get_max_bound() - mesh.get_min_bound()), center=mesh.get_center())o3d.visualization.draw_geometries([mesh])print('voxelization')voxel_grid = o3d.geometry.VoxelGrid.create_from_triangle_mesh(mesh, voxel_size=0.05)o3d.visualization.draw_geometries([voxel_grid])

input

voxelization

从点云中生成

也能够使用create_from_point_cloud函数从点云中生成体素网格。如果点云中至少有一个点在体素网格内，则该网格被占用。颜色表示的是该体素中点的平均值。参数voxel_size用来定义网格分辨率。

print('input')N = 2000pcd = o3dtut.get_armadillo_mesh().sample_points_poisson_disk(N)# fit to unit cubepcd.scale(1 / np.max(pcd.get_max_bound() - pcd.get_min_bound()), center=pcd.get_center())pcd.colors = o3d.utility.Vector3dVector(np.random.uniform(0,1,size=(N,3)))o3d.visualization.draw_geometries([pcd])

print('voxelization')voxel_grid = o3d.geometry.VoxelGrid.create_from_point_cloud(pcd, voxel_size=0.05)o3d.visualization.draw_geometries([voxel_grid])

input

voxelization

包含测试

体素网格也能够用来测试点是否在被占用的网格内。方法check_if_included接受一个(n,3)数组作为输入，返回一个bool类型的数组。

queries = np.asarray(pcd.points)output = voxel_grid.check_if_included(o3d.utility.Vector3dVector(queries))print(output[:10])

体素雕刻

create_from_point_cloud和create_from_triangle_mesh方法只能够在几何体的表面创造体素网格。然而从大量的深度图或者轮廓中雕刻一个体素网格是有可能的。Open3d提供了carve_depth_map和 carve_silhouette方法用于体素雕刻。

下面的代码展示了使用方法，首先从一个几何形状中得到 depthmaps ，之后使用 depthmaps 去雕刻出稠密的体素网格。最后的结果是一个给定形状的填充的体素网格。

def xyz_spherical(xyz): x = xyz[0] y = xyz[1] z = xyz[2] r = np.sqrt(x * x + y * y + z * z) r_x = np.arccos(y / r) r_y = np.arctan2(z, x) return [r, r_x, r_y]

def get_rotation_matrix(r_x, r_y): rot_x = np.asarray([[1, 0, 0], [0, np.cos(r_x), -np.sin(r_x)], [0, np.sin(r_x), np.cos(r_x)]]) rot_y = np.asarray([[np.cos(r_y), 0, np.sin(r_y)], [0, 1, 0], [-np.sin(r_y), 0, np.cos(r_y)]]) return rot_y.dot(rot_x)

def get_extrinsic(xyz): rvec = xyz_spherical(xyz) r = get_rotation_matrix(rvec[1], rvec[2]) t = np.asarray([0, 0, 2]).transpose() trans = np.eye(4) trans[:3, :3] = r trans[:3, 3] = t return trans

def preprocess(model): min_bound = model.get_min_bound() max_bound = model.get_max_bound() center = min_bound + (max_bound - min_bound) / 2.0 scale = np.linalg.norm(max_bound - min_bound) / 2.0 vertices = np.asarray(model.vertices) vertices -= center model.vertices = o3d.utility.Vector3dVector(vertices / scale) return model

def voxel_carving(mesh, output_filename, camera_path, cubic_size, voxel_resolution, w=300, h=300, use_depth=True, surface_method='pointcloud'): mesh.compute_vertex_normals() camera_sphere = o3d.io.read_triangle_mesh(camera_path)

# setup dense voxel grid voxel_carving = o3d.geometry.VoxelGrid.create_dense( width=cubic_size, height=cubic_size, depth=cubic_size, voxel_size=cubic_size / voxel_resolution, origin=[-cubic_size / 2.0, -cubic_size / 2.0, -cubic_size / 2.0])

# rescale geometry camera_sphere = preprocess(camera_sphere) mesh = preprocess(mesh)

# setup visualizer to render depthmaps vis = o3d.visualization.Visualizer() vis.create_window(width=w, height=h, visible=False) vis.add_geometry(mesh) vis.get_render_option().mesh_show_back_face = True ctr = vis.get_view_control() param = ctr.convert_to_pinhole_camera_parameters()

# carve voxel grid pcd_agg = o3d.geometry.PointCloud() centers_pts = np.zeros((len(camera_sphere.vertices), 3)) for cid, xyz in enumerate(camera_sphere.vertices): # get new camera pose trans = get_extrinsic(xyz) param.extrinsic = trans c = np.linalg.inv(trans).dot(np.asarray([0, 0, 0, 1]).transpose()) centers_pts[cid, :] = c[:3] ctr.convert_from_pinhole_camera_parameters(param)

# capture depth image and make a point cloud vis.poll_events() vis.update_renderer() depth = vis.capture_depth_float_buffer(False) pcd_agg += o3d.geometry.PointCloud.create_from_depth_image( o3d.geometry.Image(depth), param.intrinsic, param.extrinsic, depth_scale=1)

# depth map carving method if use_depth: voxel_carving.carve_depth_map(o3d.geometry.Image(depth), param) else: voxel_carving.carve_silhouette(o3d.geometry.Image(depth), param) print("Carve view %03d/%03d" % (cid + 1, len(camera_sphere.vertices))) vis.destroy_window()

# add voxel grid survace print('Surface voxel grid from %s' % surface_method) if surface_method == 'pointcloud': voxel_surface = o3d.geometry.VoxelGrid.create_from_point_cloud_within_bounds( pcd_agg, voxel_size=cubic_size / voxel_resolution, min_bound=(-cubic_size / 2, -cubic_size / 2, -cubic_size / 2), max_bound=(cubic_size / 2, cubic_size / 2, cubic_size / 2)) elif surface_method == 'mesh': voxel_surface = o3d.geometry.VoxelGrid.create_from_triangle_mesh_within_bounds( mesh, voxel_size=cubic_size / voxel_resolution, min_bound=(-cubic_size / 2, -cubic_size / 2, -cubic_size / 2), max_bound=(cubic_size / 2, cubic_size / 2, cubic_size / 2)) else: raise Exception('invalid surface method') voxel_carving_surface = voxel_surface + voxel_carving

return voxel_carving_surface, voxel_carving, voxel_surface

mesh = o3dtut.get_armadillo_mesh()

output_filename = os.path.abspath("../../TestData/voxelized.ply")camera_path = os.path.abspath("../../TestData/sphere.ply")visualization = Truecubic_size = 2.0voxel_resolution = 128.0

voxel_grid, voxel_carving, voxel_surface = voxel_carving( mesh, output_filename, camera_path, cubic_size, voxel_resolution)

Carve view 001/642

Carve view 002/642

Carve view 003/642

Carve view 004/642

…

Carve view 642/642

Surface voxel grid from pointcloud

print("surface voxels")print(voxel_surface)o3d.visualization.draw_geometries([voxel_surface])

print("carved voxels")print(voxel_carving)o3d.visualization.draw_geometries([voxel_carving])

print("combined voxels (carved + surface)")print(voxel_grid)o3d.visualization.draw_geometries([voxel_grid])

surface voxels