基于距离变换的稠密点云配准算法

Dense Point Cloud Registration Method based on Distance Transform

摘要:

点云配准在计算机视觉领域有着关键作用，涵盖三维重建、目标识别及机器人导航等多个领域。当前传感器技术的进步使得稠密点云数据的获取更为容易，模型几何细节也更加精确。然而，稠密点云的处理面临诸多挑战，如计算复杂、效率低下等。本文针对此提出一种基于距离变换的稠密点云配准算法。该算法首先通过距离变换技术对源点云和目标点云进行投影，获取不同视角下的距离变换图。接着以两图差异作为目标函数，利用梯度下降算法优化点云间的变换关系，经多次迭代实现准确对齐。在 Stanford 3D Scanning 数据集上的实验显示，本算法在处理稠密点云时精度高、速度快，配准数十万点云仅需约 2 秒，误差小于 0.3°和 0.5mm，具有广阔的应用前景。

Abstract:

Point cloud registration plays a crucial role in the field of computer vision, covering multiple areas such as 3D reconstruction, target recognition, and robot navigation. With the advancement of sensor technology, the acquisition of dense point cloud data has become easier, and the geometric details of models are more precise. However, the processing of dense point clouds faces many challenges, such as high computational complexity and low efficiency. To address this, this paper proposes a dense point cloud registration algorithm based on distance transformation. The algorithm first projects the source point cloud and the target point cloud through distance transformation technology to obtain distance transformation maps from different perspectives. Then, taking the difference between the two maps as the objective function, the gradient descent algorithm is used to optimize the transformation relationship between point clouds. After multiple iterations, accurate alignment is achieved. Experiments on the Stanford 3D Scanning dataset show that this algorithm has high precision and fast speed when processing dense point clouds. Registering hundreds of thousands of point clouds only takes about 2 seconds, and the error is less than 0.3° and 0.5mm. It has broad application prospects.