CloudCompare标注数据集自动化处理实战：从点云预处理到批量导出

1次阅读

共计 3476 个字符，预计需要花费 9 分钟才能阅读完成。

在使用 CloudCompare 进行点云标注时，手动处理方式存在几个明显的瓶颈：

坐标系转换漂移问题 ：手动进行坐标系转换时，由于操作误差和软件间的坐标系差异，容易导致最终标注位置出现偏差。
多文件批次管理困难 ：面对大规模数据集时，手动一个个处理文件效率极低，且容易遗漏或重复处理某些文件。
标注结果格式不统一 ：不同人员或不同批次的标注结果可能存在格式差异，给后续的数据分析和模型训练带来麻烦。

使用 Python 的 open3d 库可以方便地实现点云数据的批量加载。以下是关键代码片段：

import open3d as o3d
from pathlib import Path

def load_pointclouds(folder_path):
    pcd_files = Path(folder_path).glob('*.pcd')
    pointclouds = []
    for pcd_file in pcd_files:
        pcd = o3d.io.read_point_cloud(str(pcd_file))
        if pcd is not None:
            pointclouds.append(pcd)
    return pointclouds

设计基于 JSON Schema 的标注元数据校验系统，确保标注结果的格式统一：

import jsonschema

annotation_schema = {
    "type": "object",
    "properties": {"object_id": {"type": "string"},
        "bbox": {
            "type": "object",
            "properties": {"center": {"type": "array", "items": {"type": "number"}, "minItems": 3, "maxItems": 3},
                "size": {"type": "array", "items": {"type": "number"}, "minItems": 3, "maxItems": 3},
                "rotation": {"type": "array", "items": {"type": "number"}, "minItems": 4, "maxItems": 4}
            },
            "required": ["center", "size", "rotation"]
        }
    },
    "required": ["object_id", "bbox"]
}

def validate_annotation(annotation):
    try:
        jsonschema.validate(instance=annotation, schema=annotation_schema)
        return True
    except jsonschema.exceptions.ValidationError:
        return False

实现多线程安全的点云变换流水线，提高处理效率：

from concurrent.futures import ThreadPoolExecutor
import numpy as np

def transform_pointcloud(pcd, transform_matrix):
    points = np.asarray(pcd.points)
    points = np.hstack([points, np.ones((points.shape[0], 1))])
    transformed_points = np.dot(points, transform_matrix.T)[:, :3]
    pcd.points = o3d.utility.Vector3dVector(transformed_points)
    return pcd

def process_batch(pcd_files, transform_matrix, max_workers=4):
    with ThreadPoolExecutor(max_workers=max_workers) as executor:
        results = list(executor.map(lambda f: transform_pointcloud(o3d.io.read_point_cloud(str(f)), transform_matrix),
            pcd_files
        ))
    return results

包含四元数转旋转矩阵的数值稳定性处理：

def quaternion_to_matrix(quat):
    """将四元数转换为旋转矩阵，处理数值稳定性问题"""
    q = np.array(quat, dtype=np.float64)
    q = q / np.linalg.norm(q)  # 归一化处理

    x, y, z, w = q
    xx = x * x
    yy = y * y
    zz = z * z
    xy = x * y
    xz = x * z
    yz = y * z
    wx = w * x
    wy = w * y
    wz = w * z

    matrix = np.array([[1 - 2*(yy + zz),     2*(xy - wz),     2*(xz + wy)],
        [2*(xy + wz), 1 - 2*(xx + zz),     2*(yz - wx)],
        [2*(xz - wy),     2*(yz + wx), 1 - 2*(xx + yy)]
    ])
    return matrix

处理点云与 2D 框的投影关系：

def export_to_kitti(pcd, annotations, calib_matrix, output_path):
    """将标注结果导出为 KITTI 格式"""
    with open(output_path, 'w') as f:
        for ann in annotations:
            # 将 3D 框投影到图像平面
            bbox_3d = ann['bbox']
            center = np.array(bbox_3d['center'])
            size = np.array(bbox_3d['size'])
            rotation = quaternion_to_matrix(bbox_3d['rotation'])

            # 构建 3D 框的 8 个角点
            corners = np.array([[center[0] - size[0]/2, center[1] - size[1]/2, center[2] - size[2]/2],
                [center[0] + size[0]/2, center[1] - size[1]/2, center[2] - size[2]/2],
                # 省略其他 6 个角点...
            ])

            # 应用旋转
            corners = np.dot(corners, rotation.T)

            # 投影到图像平面
            corners_homo = np.hstack([corners, np.ones((corners.shape[0], 1))])
            img_points = np.dot(corners_homo, calib_matrix.T)
            img_points = img_points[:, :2] / img_points[:, 2:3]

            # 计算 2D 边界框
            xmin, ymin = np.min(img_points, axis=0)
            xmax, ymax = np.max(img_points, axis=0)

            # 写入 KITTI 格式
            line = f"{ann['object_id']} 0 0 0 {xmin} {ymin} {xmax} {ymax}"
            line += "".join(map(str, bbox_3d['center'])) +" "line +=" ".join(map(str, bbox_3d['size'])) +" "line +=" ".join(map(str, bbox_3d['rotation'])) +"\n"
            f.write(line)

我们测试了处理 10GB 点云数据时，单线程和多线程模式下的时间消耗：