Local Feature Descriptor Construction Technique Based on Point Pair Features and Spatial Features and Its Quantitative Evaluation

Fujing Zhou,Bao Zhao

DOI: https://doi.org/10.1109/nnice61279.2024.10498236

2024-01-01

Abstract:Three-dimensional local feature descriptors are essential tools for computer vision. This study proposes a novel method to integrate geometric information based on point pair features (PPFs) and spatial features based on the polar coordinate system to process three-dimensional point cloud data. The research constructs a system of twenty-one local feature descriptors.We perform thorough experiments on four benchmark datasets and systematically examine the descriptive capacities and resilience of these descriptors.We confirm the superior performance of the twenty-one feature descriptors. The experimental results show significant differences in the performance of the descriptors based on the three geometric features. We highlight the outstanding robustness of descriptors with complex spatial and geometric features in low-quality scenarios. We propose three descriptors, namely f r (α 1 α 2 α 3 ), f r (α 1 α 3 ), and f r (α 1 α 3 ), that outperform existing descriptors. This study provides substantial guidance and insights for the in-depth application and further improvement of three-dimensional local feature descriptors.