三维重建技术是计算机视觉领域研究的重点，广泛应用于自动驾驶、逆向工程、文物恢复、观演空间展示等领域。目前，对单物体的重建效果显著，但大场景信息复杂、特征杂乱，现有针对大场景的特征匹配算法和点云配准算法在计算效率和精度上仍存在一定的局限性。因此，本文基于RGB-D图像序列，首先采用了一种结合非极大值抑制的ORB特征提取方法，提出了基于KD树和优先队列相结合的匹配方式，然后构建了融合多元信息的关键帧筛选机制，实现了对局部场景稠密点云的实时生成。其次，提出了一种基于双重阈值约束点云精配准方法，在点云法向量夹角阈值约束的基础上，通过自适应距离阈值约束实现ICP算法中最近邻点对的搜索。最后，在真实大场景中进行实验分析，验证了本文算法的有效性。

Three-dimensional reconstruction technology is the focus of research in the field of computer vision, which is widely used in automatic driving, reverse engineering, cultural relics restoration, performance space display and other fields. At present, the reconstruction effect of single object is remarkable, but the information of large scene is complex and the features are messy. The existing feature matching algorithms and point cloud registration algorithms for large scene still have some limitations in computational efficiency and accuracy. Therefore, based on RGB-D image sequence, this paper firstly adopts an ORB feature extraction method combined with non-maximum suppression, and proposes a matching method based on KD tree and priority queue, and then constructs a key frame filtering mechanism based on multivariate information, which realizes the real-time generation of dense point clouds in local scenes. Secondly, a point cloud fine-registration method based on double threshold constraint is proposed. Based on the threshold constraint of point cloud normal vector angle, the nearest neighbor pair in ICP algorithm is searched by adaptive distance threshold constraint. Finally, experiments are carried out in real large scenes to verify the effectiveness of the proposed algorithm.