Evolutionary Multitasking With Self-Learning and Multi-Scale Transfer for Point Cloud Registration

Point cloud registration is a fundamental research topic in the field of computer vision, where the goal is to find a rotation matrix as well as a translation vector to align a pair of point clouds. Generally, point cloud registration is modeled with either point to point distance or point to plane distance, and compromises such as linearization and approximation are made to obtain mathematical solution, which are against registration accuracy. Fortunately, these two metrics are not in conflict and point cloud registration accuracy can be benefited from both of them if they are combined properly. Therefore, this paper models point cloud registration with point to point distance and improved point to plane distance, and an adaptive residual function is adopted to reduce negative effect caused by noise and outliers. Evolutionary algorithms are well-known for their global search ability even if the optimized function lacks mathematical properties such as continuity and differentiability, and evolutionary multitask optimization is proposed recently to enhance performance. Inspired by ideas from evolutionary multitasking community, this paper solves registration problems in a multitasking setting. Two new strategies, namely self-learning and multi-scale knowledge transfer, are designed to improve search ability. The self-learning strategy is introduced to estimate potential optimal solution of each component task, and differences between tasks can then be calculated to guide knowledge transfer. The multi-scale knowledge transfer strategy takes advantage of multiple differences calculated by different ratios of elite individuals to achieve positive knowledge transfer. Comprehensive experiments conducted on point clouds from various datasets have demonstrated the effectiveness of the proposed method.