NDT-Transformer: Large-Scale 3D Point Cloud Localisation using the Normal Distribution Transform Representation

Abstract

3D point cloud-based place recognition is highly demanded by autonomous driving in GPS-challenged environments and serves as an essential component (i.e. loopclosure detection) in lidar-based SLAM systems. This paper proposes a novel approach, named NDT-Transformer, for realtime and large-scale place recognition using 3D point clouds. Specifically, a 3D Normal Distribution Transform (NDT) representation is employed to condense the raw, dense 3D point cloud as probabilistic distributions (NDT cells) to provide the geometrical shape description. Then a novel NDT-Transformer network learns a global descriptor from a set of 3D NDT cell representations. Benefiting from the NDT representation and NDT-Transformer network, the learned global descriptors are enriched with both geometrical and contextual information. Finally, descriptor retrieval is achieved using a query-database for place recognition. Compared to the state-of-the-art methods, the proposed approach achieves an improvement of 7.52% on average top 1 recall and 2.73% on average top 1% recall on the Oxford Robotcar benchmark. †Authors contribute equally.This project is funded by EPSRC FAIR-SPACE Hub (EP/R026092/1), Royal Society RGS/R2/202432 and EU Horizon 2020 ILIAD (No. 732737).

3D ToF LiDAR in Mobile Robotics: A Review

“…The following work LPD-Net [49] employs a classical DGCNN [85]-like network to enhance the feature descriptor by KNN-based aggregation in both feature space and Cartesian space. Our recent work [76] proposed to employ a 3D Normal Distribution Transform (NDT) representation to condense the raw, dense 3D point cloud as probabilistic distributions (NDT cells) to provide the geometrical shape description, so as to achieve real-time and large-scale place recognition using 3D point clouds.…”

“…Fig.8. Examples of metric-based (upper)[37] and one-shot global (lower)[76] robot localization based on 3D LiDAR.…”

CGS-Net: Aggregating Colour, Geometry and Semantic Features for Large-Scale Indoor Place Recognition

“…This inspired a series of end-to-end visual place recognition networks [5], [6], [7], [8], [9]. To utilise 3-D information, PointNetVLAD [10] and its successors [11], [12], [13], [14], [15] use point clouds as inputs and achieve very high average recall rates in outdoor environments with 3-D features aggregated Fig. 1.…”

TransLoc3D : Point Cloud based Large-scale Place Recognition using Adaptive Receptive Fields

“…However, compared with point clouds from LiDAR sensors, images from RGB cameras are more sensitive to change of illumination caused by different environments, which may lead to significant performance degradation under certain circumstances [46]. To alleviate this problem, more and more works [46,55,41,27,18,57] began to focus on 3D point cloud based place recognition due to the inherent invariance of point clouds.…”

“…Given a query point cloud ("Query" at the bottom left), the purpose is to find the closest match from the prebuilt database and return its corresponding place. Existing methods can be roughly categorized into two main streams according to the data representations -point cloud based methods [46,55,53,27,41,57] and sparse voxel based methods [18,19]. PointNetVLAD [46] is the pioneer of point-based methods.…”