[Paper Review] NDD: A 3D Point Cloud Descriptor Based on Normal Distribution for Loop Closure Detection

Ruihao Zhou, Li He+3 more|arXiv (Cornell University)|Sep 26, 2022

Robotics and Sensor-Based Localization4 citations

TL;DR

This paper proposes Normal Distribution Descriptor (NDD), a novel 3D point cloud global descriptor for LiDAR-based loop closure detection that encodes both probability density score and entropy of point clouds. By leveraging correlation coefficient similarity and a fast rotation alignment via row vector transformation, NDD achieves state-of-the-art accuracy and efficiency, outperforming existing methods on KITTI and Mulran datasets with a 0.896 F1 score and 72ms total time cost.

ABSTRACT

Loop closure detection is a key technology for long-term robot navigation in complex environments. In this paper, we present a global descriptor, named Normal Distribution Descriptor (NDD), for 3D point cloud loop closure detection. The descriptor encodes both the probability density score and entropy of a point cloud as the descriptor. We also propose a fast rotation alignment process and use correlation coefficient as the similarity between descriptors. Experimental results show that our approach outperforms the state-of-the-art point cloud descriptors in both accuracy and efficency. The source code is available and can be integrated into existing LiDAR odometry and mapping (LOAM) systems.

Motivation & Objective

• To address the limitations of existing global descriptors in handling orientation sensitivity and computational inefficiency in 3D LiDAR loop closure detection.
• To develop a hand-crafted global descriptor that captures both local distribution and structural properties of point clouds without requiring deep learning or complex clustering.
• To improve matching accuracy and retrieval speed by replacing cosine similarity with correlation coefficient and introducing a fast rotation alignment method.
• To achieve high performance on standard benchmarks while maintaining low computational cost for real-time deployment.

Proposed method

• NDD encodes a 3D point cloud using two-scale statistics: the probability density score (P) and entropy (E), derived from a normal distribution model fitted to local point cloud neighborhoods.
• The method transforms the full descriptor into a 1×Ns row vector to enable fast similarity computation via correlation coefficient, replacing traditional cosine similarity.
• A novel fast rotation alignment strategy is proposed by comparing row vectors after column-wise shifting, significantly reducing alignment time compared to matrix-based comparisons.
• The descriptor is generated using a grid-based projection similar to Scan Context, with a 20×60 resolution, and uses height-based encoding for robustness to scale and resolution changes.
• A key-frame searching strategy is employed to accelerate retrieval, reducing average time cost by 99% compared to full descriptor comparison.
• The correlation coefficient is used as the similarity metric between descriptors, which is shown to be more effective than cosine similarity in matching performance.

Figure 1: Approach overview. First, one point cloud is divided into cells and one cell is represented by its probability density score and entropy. Then, A vector named searching key is compressed from full descriptor and used to create a KD-tree for retrieving the nearest neighbors. Finally, the re

Experimental results

Research questions

• RQ1 Can a hand-crafted global descriptor based on statistical properties of point clouds outperform existing state-of-the-art methods in loop closure detection accuracy and efficiency?
• RQ2 How does combining probability density score and entropy improve descriptor robustness to viewpoint and noise variations?
• RQ3 Does replacing cosine similarity with correlation coefficient lead to measurable improvements in matching performance on real-world LiDAR datasets?
• RQ4 To what extent does the proposed fast rotation alignment method reduce computational cost without sacrificing accuracy?
• RQ5 Can the descriptor be efficiently integrated into existing LiDAR odometry and mapping (LOAM) systems?

Key findings

• The proposed NDD descriptor achieved a 0.896 F1 score on the KITTI 08 dataset, outperforming all one-feature baselines and the next-best method (ISC) by 3.6 percentage points.
• The combination of probability density score and entropy (P+E) in NDD yielded a 13.1% higher F1 score than the best single-feature method (P), demonstrating the effectiveness of multi-scale statistical encoding.
• Using correlation coefficient similarity improved F1 by 8.2% and EP by 5.4% compared to cosine similarity, confirming its superiority in descriptor matching.
• The total time cost of NDD was 72ms on KITTI 08, significantly lower than IRIS (5533ms) and M2DP (180ms), and comparable to ISC (76ms) but with much higher accuracy.
• The row vector-based rotation alignment reduced alignment time to 9.78ms, a 53% and 56% improvement over matrix-based and binarization-based methods, respectively.
• On the Mulran Sejong 01 dataset, key-frame searching reduced retrieval time from 1567ms to 35ms with only a 0.4% drop in F1 score, proving its practical efficiency.

Figure 2: Cell division of a BEV representation