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 3647 个字符，预计需要花费 10 分钟才能阅读完成。

在处理点云标注数据时，手动操作不仅效率低下，还容易出错。经过实际项目验证，我们发现以下三大瓶颈问题：

坐标系转换漂移 ：手动在 CloudCompare 中逐个文件应用变换矩阵时，由于界面操作误差，容易导致累计误差增大
多文件批次管理困难 ：当处理数百个 PCD 文件时，人工管理文件命名、版本和转换状态极易混乱
标注结果格式不统一 ：不同标注人员导出的结果可能采用不同坐标系或单位制，给后续算法训练带来麻烦

使用 Python 的 open3d 库实现高效的点云文件批量读取：

import open3d as o3d
from pathlib import Path

def load_pointclouds(folder_path):
    """
    批量加载点云文件（支持 PCD/PLY 格式）:param folder_path: 包含点云文件的目录路径
    :return: 点云数据字典 {filename: pointcloud}
    """
    pc_dict = {}
    for f in Path(folder_path).glob('*.[pP][cC][dD]'):
        pc = o3d.io.read_point_cloud(str(f))
        pc_dict[f.stem] = pc
    return pc_dict

基于 JSON Schema 设计严格的标注规范检查：

from jsonschema import validate

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": "number"}
            },
            "required": ["center", "size", "rotation"]
        }
    },
    "required": ["object_id", "bbox"]
}

def validate_annotation(annotation):
    try:
        validate(instance=annotation, schema=annotation_schema)
        return True
    except Exception as e:
        print(f"Invalid annotation: {str(e)}")
        return False

采用线程池处理大规模点云转换任务：

from concurrent.futures import ThreadPoolExecutor
import numpy as np

class PointCloudProcessor:
    def __init__(self, max_workers=4):
        self.executor = ThreadPoolExecutor(max_workers=max_workers)

    def apply_transform(self, pc, transform_matrix):
        """线程安全的点云变换"""
        points = np.asarray(pc.points)
        # 添加齐次坐标
        homogeneous = np.column_stack([points, np.ones(points.shape[0])])
        transformed = (transform_matrix @ homogeneous.T).T[:, :3]
        pc.points = o3d.utility.Vector3dVector(transformed)
        return pc

    def batch_process(self, pc_dict, transforms):
        futures = []
        results = {}
        for name, pc in pc_dict.items():
            if name in transforms:
                future = self.executor.submit(self.apply_transform, pc.copy(), transforms[name])
                futures.append((name, future))

        for name, future in futures:
            results[name] = future.result()
        return results

def quaternion_to_matrix(quat):
    """四元数转旋转矩阵（带数值稳定性检查）"""
    q = np.array(quat)
    q_norm = np.linalg.norm(q)
    if abs(q_norm - 1.0) > 1e-6:
        q = q / q_norm  # 归一化处理

    qx, qy, qz, qw = q
    return np.array([[1 - 2*qy**2 - 2*qz**2, 2*qx*qy - 2*qz*qw, 2*qx*qz + 2*qy*qw],
        [2*qx*qy + 2*qz*qw, 1 - 2*qx**2 - 2*qz**2, 2*qy*qz - 2*qx*qw],
        [2*qx*qz - 2*qy*qw, 2*qy*qz + 2*qx*qw, 1 - 2*qx**2 - 2*qy**2]
    ])

def create_transform_matrix(rotation, translation):
    """创建 4x4 齐次变换矩阵"""
    mat = np.eye(4)
    mat[:3, :3] = quaternion_to_matrix(rotation)
    mat[:3, 3] = translation
    return mat

def export_to_kitti(pc, annotations, calib):
    """
    导出为 KITTI 格式标注文件
    :param pc: 点云对象
    :param annotations: 标注信息列表
    :param calib: 相机标定参数
    """
    lines = []
    points = np.asarray(pc.points)

    # 将点云投影到图像平面
    pts_rect = calib['R0_rect'] @ (calib['Tr_velo_to_cam'] @ 
                                  np.vstack([points.T, np.ones(points.shape[0])]))
    pts_img = (calib['P2'] @ pts_rect).T
    pts_img[:, :2] /= pts_img[:, [2]]  # 齐次坐标归一化

    for anno in annotations:
        bbox = anno['bbox']
        # 计算 2D 包围框（简化版）x_min = min(pts_img[:, 0])
        y_min = min(pts_img[:, 1])
        x_max = max(pts_img[:, 0])
        y_max = max(pts_img[:, 1])

        line = f"{anno['object_id']} 0 0 0 {x_min} {y_min} {x_max} {y_max}" \
               f"{bbox['size'][0]} {bbox['size'][1]} {bbox['size'][2]}" \
               f"{bbox['center'][0]} {bbox['center'][1]} {bbox['center'][2]}" \
               f"{bbox['rotation']}\n"
        lines.append(line)
    return lines

测试环境：AMD Ryzen 7 5800H, 32GB RAM