DeepCLR: Correspondence-Less Architecture for Deep End-to-End Point Cloud Registration

Horn, Markus; Engel, Nico; Belagiannis, Vasileios; Buchholz, Michael; Dietmayer, Klaus

doi:10.1109/itsc45102.2020.9294279

“…The evaluation resulted in Et = 5.4% and Er = 0.0154 deg/m and is currently ranked 121/134. This is remarkable as our model consists of only 3.56% of the number of parameters of DeepCLR [1] which is ranked 118/134 with Et = 4.19% and Er = 0.0087 deg/m (numbers are as of January 25, 2021, the date of submission of this work).…”

Section: Resultsmentioning

confidence: 98%

D3dlo: Deep 3d Lidar Odometry

Adis

¹

,

Horst

²

,

Wien

³

2021

2021 IEEE International Conference on Image Processing (ICIP)

4

0

View full text Add to dashboard Cite

LiDAR odometry (LO) describes the task of finding an alignment of subsequent LiDAR point clouds. This alignment can be used to estimate the motion of the platform where the LiDAR sensor is mounted on. Currently, on the well-known KITTI Vision Benchmark Suite state-of-the-art algorithms are non-learning approaches. We propose a network architecture that learns LO by directly processing 3D point clouds. It is trained on the KITTI dataset in an end-to-end manner without the necessity of pre-defining corresponding pairs of points. An evaluation on the KITTI Vision Benchmark Suite shows similar performance to a previously published work, DeepCLR [1], even though our model uses only around 3.56% of the number of network parameters thereof. Furthermore, a plane point extraction is applied which leads to a marginal performance decrease while simultaneously reducing the input size by up to 50%.

show abstract

“…Based on the findings of previous work, which we will summarize in section 2, the main part of this work is dedicated to the introduction of our proposed model in section 3. It is similar to [1] and performs comparably while using only about 3.56% the number of trainable parameters thereof. In sections 4 and 5 the training and evaluation of the model are introduced.…”

Section: Introductionmentioning

confidence: 87%

“…Unfortunately, in pure LO no initial guess of the target transformation is given. This can lead to large inaccuracies when applying ICP at high velocities of the moving platform due to the large relative displacement of subsequent point clouds [1]. In recent years LOAM [13] has been considered as a state-of-the-art approach for LO and is also ranked as the best LO-method on the KITTI Vision Benchmark Suite.…”

Section: Related Workmentioning

confidence: 99%

D3DLO: Deep 3D LiDAR Odometry

Adis¹,

Horst²,

Wien³

2021

Preprint

0

View full text Add to dashboard Cite

LiDAR odometry (LO) describes the task of finding an alignment of subsequent LiDAR point clouds. This alignment can be used to estimate the motion of the platform where the LiDAR sensor is mounted on. Currently, on the well-known KITTI Vision Benchmark Suite state-of-the-art algorithms are non-learning approaches. We propose a network architecture that learns LO by directly processing 3D point clouds. It is trained on the KITTI dataset in an end-to-end manner without the necessity of pre-defining corresponding pairs of points. An evaluation on the KITTI Vision Benchmark Suite shows similar performance to a previously published work, DeepCLR [1], even though our model uses only around 3.56% of the number of network parameters thereof. Furthermore, a plane point extraction is applied which leads to a marginal performance decrease while simultaneously reducing the input size by up to 50%.

show abstract

“…This makes them suitable for many sensor types such as cameras, lidars, radars, or IMUs. Although per-sensor ego-motion is not directly available from most sensors, algorithms used for registration and mapping are effective tools to estimate it [9]- [11]. Since we want to offer a universal calibration tool that can be deployed on many different sensor setups, our work is focused on motion-based extrinsic calibration.…”

Section: Globalmentioning

confidence: 99%

“…For estimating the required per-sensor ego-motion, registration algorithms [4], [11] or online capable SLAM algorithms like LOAM for lidars [9] or OpenVSLAM for cameras [10] can be used. Additionally, SLAM algorithms are often directly used to estimate the sensor calibration by registering the individual maps of multiple per-sensor SLAM algorithms [14]- [16].…”

Section: Related Workmentioning

confidence: 99%

Online Extrinsic Calibration based on Per-Sensor Ego-Motion Using Dual Quaternions

Horn,

Wodtko,

Buchholz

et al. 2021

Preprint

Self Cite

0

View full text Add to dashboard Cite

In this work, we propose an approach for extrinsic sensor calibration from per-sensor ego-motion estimates. Our problem formulation is based on dual quaternions, enabling two different online capable solving approaches. We provide a certifiable globally optimal and a fast local approach along with a method to verify the globality of the local approach. Additionally, means for integrating previous knowledge, for example, a common ground plane for planar sensor motion, are described. Our algorithms are evaluated on simulated data and on a publicly available dataset containing RGB-D camera images. Further, our online calibration approach is tested on the KITTI odometry dataset, which provides data of a lidar and two stereo camera systems mounted on a vehicle. Our evaluation confirms the short run time, state-of-the-art accuracy, as well as online capability of our approach while retaining the global optimality of the solution at any time.

show abstract

DeepCLR: Correspondence-Less Architecture for Deep End-to-End Point Cloud Registration

Cited by 21 publications

References 20 publications

D3dlo: Deep 3d Lidar Odometry

D3dlo: Deep 3d Lidar Odometry

D3DLO: Deep 3D LiDAR Odometry

Online Extrinsic Calibration based on Per-Sensor Ego-Motion Using Dual Quaternions

Contact Info

Product

Resources

About