多头门控多层感知机运动分解非刚性点云配准算法
Multi-head Gated Multilayer Perceptron Motion Decomposition Non-rigid Point Cloud Registration Algorithm
-
摘要: 非刚性物体运动的复杂性和不确定性, 使非刚性点云配准成为一个极具挑战性的问题. 现有的非刚性点云配准算法配准结果存在大量外点, 面对噪声输入时, 鲁棒性不高. 为解决以上问题, 提出了一种多头门控多层感知机运动分解非刚性点云配准算法. 首先, 对三维点云中的点进行正弦编码, 输入到多层级多头门控多层感知机中; 其次, 通过运动分解, 在每一层级中用多头门控单元来解决外点与噪声问题, 在最后一个层级中得到对应点的运动增量; 在这一过程中, 随着层级增加逐渐提高正弦函数的频率, 以增加函数的波动性来表示非刚性运动; 最后, 在2个数据集上与9种主流的非刚性点云配准算法用4个评价指标进行了对比实验. 该算法与其他对比算法相比端点误差与异常率2个评价指标分别降低了1.84~18.8与5.61%~30.76%, 严格正确配准率与正确配准率分别提高了2.78%~13.26%与5.43%~22.63%, 而配准时间缩短了2到50倍. 实验结果表明本文算法可以快速地得到点云分布均匀且平滑, 无外点的非刚性点云配准结果.Abstract: Non-rigid point cloud registration has been a highly challenging problem due to the complexity and uncertainty of non-rigid object motion. Existing algorithms for non-rigid point cloud registration have suffered from a significant number of outliers in the registration results and limited robustness against noise input. To address these issues, researchers proposed a novel approach called the multi-head gated multi-layer perceptron motion decomposition algorithm.In this algorithm, the points in the 3D point cloud were first encoded using sine functions and then inputted into a multi-level multi-head gated multi-layer perceptron. By decomposing the motion, multi-head gated units were utilized at each level to handle outliers and noise, ultimately obtaining the motion increments for corresponding points in the final level. Throughout this process, the frequency of the sine function gradually increased with each level to capture the fluctuation patterns of non-rigid motion.To evaluate the effectiveness of the proposed algorithm, experiments were conducted on two datasets and compared against nine mainstream non-rigid point cloud registration methods using four evaluation metrics. The results demonstrated that the proposed algorithm achieved a reduction in endpoint error and outlier rate ranging from 1.84% to 18.8% and 5.61% to 30.76%, respectively. Moreover, it exhibited an improvement in the strict correctness registration rate and correct registration rate by 2.78% to 13.26% and 5.43% to 22.63%, respectively.Additionally, the registration time was reduced by a factor of 2 to 50. These experimental findings highlight the algorithm's ability to rapidly produce non-rigid point cloud registration results with a uniform and smooth point cloud distribution, free from outliers.