Self-Supervised Point Set Local Descriptors for Point Cloud Registration

In the vehicle pose estimation task based on roadside Lidar in cooperative perception, the measurement distance, angle, and laser resolution directly affect the quality of the target point cloud. For incomplete and sparse point clouds, current methods are either less accurate in correspondences solved by local descriptors or not robust enough due to the reduction of effective boundary points. In response to the above weakness, this paper proposed a registration algorithm Environment Constraint Principal Component-Iterative Closest Point (ECPC-ICP), which integrated road information constraints. The road normal feature was extracted, and the principal component of the vehicle point cloud matrix under the road normal constraint was calculated as the initial pose result. Then, an accurate 6D pose was obtained through point-to-point ICP registration. According to the measurement characteristics of the roadside Lidars, this paper defined the point cloud sparseness description. The existing algorithms were tested on point cloud data with different sparseness. The simulated experimental results showed that the positioning MAE of ECPC-ICP was about 0.5% of the vehicle scale, the orientation MAE was about 0.26°, and the average registration success rate was 95.5%, which demonstrated an improvement in accuracy and robustness compared with current methods. In the real test environment, the positioning MAE was about 2.6% of the vehicle scale, and the average time cost was 53.19 ms, proving the accuracy and effectiveness of ECPC-ICP in practical applications.

Section: Model-based Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

ECPC-ICP: A 6D Vehicle Pose Estimation Method by Fusing the Roadside Lidar Point Cloud and Road Feature

Liu

Xiong

et al. 2021

“…For example, if the first point has 10 neighborhood points with the shape of (1,10,3), while the second point has 20 neighborhood points with the shape of (1,20,3), the network cannot stack these two points for learning. However, if both of the shapes of the two points is (1,20,3), the network can stack the two points into the shape of (2,20,3). Therefore, in this paper, the number of initial neighborhood points is fixed to k, and the mask M i j is used to remove the pseudo neighborhood points from the neighborhood since these points are not conducive to the network learning of the local region.…”

Section: Masking Mechanismmentioning

confidence: 99%

“…With the rapid development of three dimensional (3D) sensing technologies, using deep learning to understand and analyze point clouds is becoming one of the important research topics [1][2][3]. As the output of 3D sensor, point cloud is composed of much number of points in 3D space.…”

Section: Introductionmentioning

confidence: 99%

DNet: Dynamic Neighborhood Feature Learning in Point Cloud

Tian

Jiang

2021

Neighborhood selection is very important for local region feature learning in point cloud learning networks. Different neighborhood selection schemes may lead to quite different results for point cloud processing tasks. The existing point cloud learning networks mainly adopt the approach of customizing the neighborhood, without considering whether the selected neighborhood is reasonable or not. To solve this problem, this paper proposes a new point cloud learning network, denoted as Dynamic neighborhood Network (DNet), to dynamically select the neighborhood and learn the features of each point. The proposed DNet has a multi-head structure which has two important modules: the Feature Enhancement Layer (FELayer) and the masking mechanism. The FELayer enhances the manifold features of the point cloud, while the masking mechanism is used to remove the neighborhood points with low contribution. The DNet can learn the manifold features and spatial geometric features of point cloud, and obtain the relationship between each point and its effective neighborhood points through the masking mechanism, so that the dynamic neighborhood features of each point can be obtained. Experimental results on three public datasets demonstrate that compared with the state-of-the-art learning networks, the proposed DNet shows better superiority and competitiveness in point cloud processing task.

“…However, many of them are prone to converging to local optima. With the advent of deep neural networks (DNNs), it has been shown [ 6 , 7 , 8 ] that PCReg methods using DNNs can achieve higher accuracy and robustness to inaccurate transformation when compared to traditional methods. The learning-based PCReg method processes unordered point clouds and extracts features through a deep learning network [ 9 , 10 , 11 ]; then, the similarity of these features is used to calculate the transformation.…”

Section: Introductionmentioning

confidence: 99%

TIF-Reg: Point Cloud Registration with Transform-Invariant Features in SE(3)

Chen

Song

et al. 2021

Three-dimensional point cloud registration (PCReg) has a wide range of applications in computer vision, 3D reconstruction and medical fields. Although numerous advances have been achieved in the field of point cloud registration in recent years, large-scale rigid transformation is a problem that most algorithms still cannot effectively handle. To solve this problem, we propose a point cloud registration method based on learning and transform-invariant features (TIF-Reg). Our algorithm includes four modules, which are the transform-invariant feature extraction module, deep feature embedding module, corresponding point generation module and decoupled singular value decomposition (SVD) module. In the transform-invariant feature extraction module, we design TIF in SE(3) (which means the 3D rigid transformation space) which contains a triangular feature and local density feature for points. It fully exploits the transformation invariance of point clouds, making the algorithm highly robust to rigid transformation. The deep feature embedding module embeds TIF into a high-dimension space using a deep neural network, further improving the expression ability of features. The corresponding point cloud is generated using an attention mechanism in the corresponding point generation module, and the final transformation for registration is calculated in the decoupled SVD module. In an experiment, we first train and evaluate the TIF-Reg method on the ModelNet40 dataset. The results show that our method keeps the root mean squared error (RMSE) of rotation within 0.5∘ and the RMSE of translation error close to 0 m, even when the rotation is up to [−180∘, 180∘] or the translation is up to [−20 m, 20 m]. We also test the generalization of our method on the TUM3D dataset using the model trained on Modelnet40. The results show that our method’s errors are close to the experimental results on Modelnet40, which verifies the good generalization ability of our method. All experiments prove that the proposed method is superior to state-of-the-art PCReg algorithms in terms of accuracy and complexity.