Iterative Overlap Attention-Aware Network with Similarity Learning for Partial Point Cloud Registration

• Published in: IEEE sensors journal 2024-04, Vol.24 (8), p.1-1
• Main Authors: Chen, Xinyu, Luo, Jiahui, Ren, Yan, Wang, Chuanyun
• Format: Article
• Language: English
• Subjects: Cloud computing; Computer vision; Deep learning; dual branches feature fusion; Estimation; Feature extraction; Image registration; Iterative methods; Machine vision; overlap factor; Point cloud compression; Point cloud registration; Registration; Robotics; Rotation; rotation invariance; Sensors; Similarity; similarity matrix; Task analysis; Three dimensional models; Three-dimensional displays
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3370994

Point cloud registration aims at accurately aligning and integrating multiple point cloud data by scanning the same object from different viewpoints into a more comprehensive 3D model or scene. It has significant applications in the field of computer vision and robotics. In recent years, with the continuous impact of machine vision technology, more and more registration methods based on deep network models have emerged. However, most deep learning-based methods perform poorly in low overlap scenarios. Therefore, this paper proposes a novel network architecture to pursue better performance with low overlapping regions. A spatial rotation feature encoder with point attention (PASR) is addressed to improve the rotation invariance of point clouds and enhance the network's perception of local feature extraction. Then this paper introduces the overlap attention prediction module (OAP) for the estimation process of point cloud overlap factors. On this basis, the cross-attention mechanism is introduced to regress the initial transformation between two input point clouds. In addition, by combining the previous overlap factors, we constructed an iterative dual-branch similarity matrix learning network (DBSML) which guides similarity estimation and further eliminates interference from non-overlapping points. Extensive experiments on ModelNet40 and our real datasets with noisy and partially overlapping point clouds show that the proposed method outperforms the traditional and mainstream learning-based methods, achieving the state-of-the-art performance. In particular, we also verify the effectiveness and superiority of the network model in coping with multiple registration task scenarios.