PointLoc: Deep Pose Regressor for LiDAR Point Cloud Localization

Abstract

In this paper, we present a novel end-to-end learning-based LiDAR sensor relocalization framework, termed PointLoc, which infers 6-DoF poses directly using only a single point cloud as input. Compared to visual sensorbased relocalization, LiDAR sensors can provide rich and robust geometric information about a scene. However, point clouds of LiDAR sensors are unordered and unstructured making it difficult to apply traditional deep learning regression models for this task. We address this issue by proposing a novel PointNet-style architecture with self-attention to efficiently estimate 6-DoF poses from 360 • LiDAR sensor frames. Extensive experiments on recently released challenging Oxford Radar RobotCar dataset and real-world robot experiments demonstrate that the proposed method can achieve accurate relocalization performance.

Slice Transformer and Self-supervised Learning for 6DoF Localization in 3D Point Cloud Maps

“…6DoF localization is an important task for self-driving vehicles [13], [14]. Whereas conventional methods for matching point cloud frames for such a task have used registration techniques [1], [15], more recent works focus on exploiting deep learning to relate an input frame to a 3D map [6], [10], [16], [17]. Among these methods, there are contributions that compress the map into a neural model and use that model as a 6DoF pose predictor for the vehicle [6], [10].…”

“…Whereas conventional methods for matching point cloud frames for such a task have used registration techniques [1], [15], more recent works focus on exploiting deep learning to relate an input frame to a 3D map [6], [10], [16], [17]. Among these methods, there are contributions that compress the map into a neural model and use that model as a 6DoF pose predictor for the vehicle [6], [10]. Using raw LiDAR frames, this prediction is particularly challenging due to the unstructured nature of the data, which conflicts with the high precision requirements of the task.…”

“…To contain the prediction latency within practical limits, the state-of-the-art methods compress the map into a neural model, and directly localize vehicles by regressing its 6DoF pose over the map [6]. For 3D point clouds, this demands high-fidelity representation learning by the model because of the unstructured nature of the data.…”

“…We establish the baseline results for 6DoF localization on our dataset with the proposed method and PointLoc [6]. We also provide benchmarking of our technique on an existing Appolo-Southbay dataset [10].…”

Sequence matching enhanced 3D place recognition using candidate rearrangement

“…Although the prior methods have achieved good results, there are still some open problems to be solved. Some works [5,6] solve place recognition as a regression problem. Given a place, these methods estimate its position directly.…”