为解决LIO-SAM （tightly-coupled lidar inertial odometry via smoothing and mapping）等基于ICP（iterative closest point）方法的传统点云配准策略和HRegNet（hierarchical registration network）等基于深度神经网络方法的新型点云配准模型均存在的算力消耗高、配准时间长等问题，通过深入研究HRegNet神经网络点云配准模型框架，提出具有轻量化、实时性特点的HKRNet（hierarchical kcpstack registration network）网络模型。首先，使用点云体素化和高斯阈值降采样联用的滤波方法，去除雷达扫描地面的大量无用点，将点的数量从13万左右降至约7万。其次，对HRegNet模型内部耗时大的K最近邻（K-nearest neighbors, KNN）点云聚类算法改进为KD树（K-dimensional tree, KD-Tree）算法，能够在保证精度的前提下提升25%的处理速度。最后，针对模型内部卷积模块内存消耗高、计算效率低的问题，使用张量分解的轻量卷积模块并提出分层奇异值分解算法，将模型压缩至原来的86.1%并节约61.2%的计算量。结果表明，HKRNet网络相对于HRegNet网络可以在微小的精度损失下，减少40%的配准时间，单次配准时间不超过84 ms，满足实时配准的使用需求。

To tackle the computational cost and registration time challenges in traditional point cloud registration methods like ICP(iterative closest point) such as LIO-SAM (tightly-coupled lidar inertial odometry via smoothing and mapping) and newer models utilizing deep neural networks such as HRegNet(hierarchical registration network), a lightweight and real-time HKRNet (hierarchical kcpstack registration network) network model was proposed. The model was developed by thoroughly studying the HRegNet neural network point cloud registration framework. Initially, a combined filtering approach involving point cloud voxelization and Gaussian threshold downsampling was used to remove redundant points from ground radar scans, reducing the point count from around 130 000 to about 70 000. Subsequently, the computationally intense KNN (K-nearest neighbors) point cloud clustering algorithm within the HRegNet model was enhanced by optimizing it to a KD-Tree (K-dimensional tree) algorithm, resulting in a 25% improvement in processing speed while upholding accuracy. Lastly, to address high memory usage and low computational efficiency of the convolutional modules in the model, a lightweight convolutional module leveraging tensor decomposition and a hierarchical singular value decomposition algorithm was introduced. This leaded to a compressed model size of 86.1% of the original and a decrease of 61.2% in computational cost. The outcomes indicate that the HKRNet network, in comparison to the HRegNet network, can reduce registration time by 40% with minimal loss of accuracy, achieving a single registration time not exceeding 84ms, thus meeting real-time registration requirements.