Narrow Results

In this paper, we propose a novel efficient surface reconstruction method from unorganized point cloud data in three-dimensional Euclidean space. The proposed method is based on the Allen–Cahn partial differential equation, with an edge indicating function to restrict the evolution. We applied the explicit Euler’s method to solve the discrete equation, and use the operator splitting technique to split the governing equation. Furthermore, we also modify the double well form to a periodic potential. Then we find that the proposed model can reconstruct the surface well even in the case of insufficient data. After selecting the appropriate parameters, we carried out various numerical experiments to demonstrate the robustness and accuracy of the proposed method. We adopt the proposed method to reconstruct the surfaces on simple, irregular and complex models, respectively, and can obtain smooth three-dimensional surfaces and visual effects. In addition, we also perform comparison tests to show the superiority of the proposed model. Statistic metrics such as the $\sigma$, $d_{max}$, $d_{\mathrm{\text{mean}}}$, CPU time, and vertex numbers are evaluated. Results show that our model performs better than the other methods in statistical metrics even use far less point cloud data, and with the faster CPU computing speed.

A 3D reconstruction method based on dynamic graph convolutional occupancy networks is proposed to address the issues of texture information loss, geometric information loss after voxelization, and lack of object completeness constraints in the process of 3D reconstruction using voxel representation in a block-wise manner. By constructing a dynamic graph structure for feature extraction, the method aims at restore 3D models with fewer holes and local details. In the feature extraction stage, local pooling is employed within each point cloud block to address the problem of nonsignificant texture feature loss. To tackle the issues of geometric constraint loss and insufficient scene semantic information caused by block-wise processing, a feature fusion method between adjacent blocks is proposed to learn richer scene semantic information and long-range dependencies between points. By learning features within and between blocks, each point retains as much geometric information as possible, mitigating the problem of geometric information loss due to voxelization. During the surface generation, interpolation is used to infer the occupancy value for each point, and the Marching Cubes algorithm is employed for three-dimensional surface reconstruction. Experimental validation on object-level (ShapeNet dataset) and scene-level (Synthetic Rooms dataset, MatterPort3D dataset for real-world scenes) datasets demonstrates the effectiveness and advancement of the proposed method.

Robust head pose estimation significantly improves the performance of applications related to face analysis in Cyber-Physical Systems (CPS) such as driving assistance and expression recognition. However, there exist two main challenges in this issue, i.e., the large pose variations and the property of inhomogeneous facial feature space. Head pose in large variations makes the distinguished facial features, such as nose or lips, invisible, especially in extreme cases. Additionally, features extracted from a head do not change in a stationary manner with respect to the head pose, which results in an inhomogeneous feature space. To deal with the above problems, we propose an end-to-end framework to estimate the head pose from a single depth image. To be specific, the PointNet network is adopted to automatically select distinguished facial feature points from visible surface of a head and to extract discriminative features. The Deep Regression Forest is utilized to handle the nonstationary property of the facial feature space and to learn the head pose distributions. Experimental results show that our proposed method achieves the state-of-the-art performance on the Biwi Kinect Head Pose Dataset, the Pandora Dataset and the ICT-3DHP Dataset.

Aiming at the problems of high complexity and low detection accuracy of single-stage three-dimensional (3D) detection method, a vehicle object detection algorithm based on the Efficient Channel Attention (ECA) mechanism is proposed. This paper provides a good solution to the problems of low object recognition accuracy and high model complexity in the field of 3D object detection. First, we voxelized the original point cloud data, taking the average coordinates and intensity values as the initial features. By entering into the Voxel Feature Encoding (VFE) layer, we can extract the features of each voxel. Then, referring to the VoxelNet model, the ECA mechanism is introduced, which reduces the complexity of the model while maintaining the good performance in the model. Finally, experiments on the widely used KITTI dataset show that the algorithm performs well, and the accuracy of the proposed ECA algorithm has reached 87.75%. Compared with the current mainstream algorithm SE-SSD of object detection, the accuracy is increased by 0.21%.

The work aims to explore a microscopic observation system of paper surface and achieve high-precision stereoscopic observation with detail characterization of paper surface morphology. Based on the DT-400E precision program-controlled three-dimensional translation stage and KEYENCE LJV-7200 two-dimensional laser scanner, the hardware parts of our own system are developed to scan and transmit point cloud data of paper surface morphology to the computer. The corresponding system software will automatically process the point cloud data acquired from the laser scanner and generate the corresponding vivid 3D model and height histogram. This system scans and characterizes four different types of paper samples, allowing the human eye to visually distinguish the differences in surface morphology as well as automatically calculate the numerical differences in paper surface morphology parameters. The results of the applicability test show that the system is highly efficient in acquiring, observing, and evaluating the topography of the paper surface. The system can not only predict the paper surface quality of printed paper, but can also be extended to the evaluation of 3D printed surfaces.

When conducting a numerical simulation of a train’s derailment and post-derailment, it is necessary to continuously observe the relative position of the wheel and rail, which is of great significance for the correct evaluation of train safety. In this paper, a non-analytic method is proposed to extend the search range and improve the accuracy of the classical semi-analytical method, i.e. the contact locus method. Based on the point cloud convex hull, a high-density wheel contact locus vertical profile is obtained by projecting the chamfer and internal zone of the flange onto the rail cutting plane. To obtain maximum compression in the normal direction and avoid singularities on both rail head sides in the Cartesian coordinate system the rail surface is interpolated with the polar spline curve. Compared with the classical method used to describe the wheel contact locus, in the proposed hybrid method, potential contact points are provided. Finally, the proposed hybrid method and the classical methods are applied to the wheel track coupling simulation. Numerical results demonstrate the high reliability and effectiveness of the proposed method.

Domain adaption is a special transfer learning method, whose source domain and target domain generally have different data distribution, but need to complete the same task. There have been many significant types of research on domain adaptation in 2D images, but in 3D data processing, domain adaptation is still in its infancy. Therefore, we design a novel domain adaptive network to complete the unsupervised point cloud classification task. Specifically, we propose a multi-scale transform module to improve the feature extractor. Besides, a spatial-awareness attention module combined with channel attention to assign weights to each node is designed to represent hierarchically scaled features. We have validated the proposed method on the PointDA-10 dataset for domain adaption classification tasks. Empirically, it shows strong performance on par or even better than state-of-the-art.

For robots in an unstructured work environment, grasping unknown objects that have neither model data nor RGB data is very important. The key to robotic autonomous grasping is not only in the judgment of object type but also in the shape of the object. We present a new grasping approach based on the basic compositions of objects. The simplification of complex objects is conducive to the description of object shape and provides effective ideas for the selection of grasping strategies. First, the depth camera is used to obtain partial 3D data of the target object. Then the 3D data are segmented and the segmented parts are simplified to a cylinder, a sphere, an ellipsoid, and a parallelepiped according to the geometric and semantic shape characteristics. The grasp pose is constrained according to the simplified shape feature and the core part of the object is used for grasping training using deep learning. The grasping model was evaluated in a simulation experiment and robot experiment, and the experiment result shows that learned grasp score using simplified constraints is more robust to gripper pose uncertainty than without simplified constraint.

The creation of interactive virtual reality (VR) applications from 3D scanned content usually includes a lot of manual and repetitive work. Our research aim is to develop agents that recognize objects to enhance the creation of interactive VR applications. We trained partition agents in our superpoint growing environment that we extended with an expert function. This expert function solves the sparse reward signal problem of the previous approaches and enables to use a variant of imitation learning and deep reinforcement learning with dense feedback. Additionally, the function allows to calculate a performance metric for the degree of imitation for different partitions. Furthermore, we introduce an environment to optimize the superpoint generation. We trained our agents with 1182 scenes of the ScanNet data set. More specifically, we trained different neural network architectures with 1170 scenes and tested their performance with 12 scenes. Our intermediate results are promising such that our partition system might be able to assist the VR application development from 3D scanned content in near future.

Neurons can be abstractly represented as skeletons due to the filament nature of neurites. With the rapid development of imaging and image analysis techniques, an increasing amount of neuron skeleton data is being produced. In some scientific studies, it is necessary to dissect the axons and dendrites, which is typically done manually and is both tedious and time-consuming. To automate this process, we have developed a method that relies solely on neuronal skeletons using Geometric Deep Learning (GDL). We demonstrate the effectiveness of this method using pyramidal neurons in mammalian brains, and the results are promising for its application in neuroscience studies.

Self-supervised learning methods are able to learn latent features from unlabeled data samples by designing pre-text tasks, and hence have attracted a great deal of interest in terms of sample-efficient learning. As a promising scheme of self-supervised learning, masked autoencoding has significantly advanced natural language processing and computer vision, but has not been introduced to point cloud yet. In this paper, we propose a novel scheme of masked autoencoders for 3D point cloud self-supervised learning, addressing the special challenges posed by point cloud, including leakage of location information and uneven information density. Concretely, we divide the input point cloud into irregular point patches and randomly mask them at a high ratio. Then, a standard Transformer-based autoencoder, with an asymmetric design and a shifting mask tokens operation, learns high-level latent features from unmasked point patches, aiming to reconstruct the masked point patches. Extensive experiments show that our approach is efficient during pre-training and generalizes well on various downstream tasks. Apart from the proposed method, we will also introduce potential directions for 3D point cloud self-supervised learning, including improvements for masked autoencoding, developments for point cloud scene understanding, etc.

Point cloud models built with large amounts of data contain more surface detail, but are a severe challenge to the processing and rendering speed of the computer. Point clouds which are too dense can cause the geometric characteristics of the measured entity to be difficult to judge, thus making it very important to simplify the original data. In this paper, the K nearest neighbor of the point cloud is first analyzed, and principal component analysis is used to estimate the normal vector of the scattered point cloud. Then, a method similar to the progressive mesh using a set of tangent planes to approximate the local surface is adopted. Finally, geometric deviation of tangent planes is estimated using the square distance tangent planes, after which the point cloud is simplified. Through verification, it was found that the method has the advantages of having a high reduction rate and the ability to maintain the geometrical features of the point cloud well.

To overcome the instability of 3D body scanning by manual operation on Kinect sensor, such as skipping frame, missing data, and time consuming, an automatic 3D scanning device using a single Kinect sensor is designed in terms of the KinectFusion algorithm. The control system of computer - Arduino controller - motor drive - stepper motor - power supply has been employed to realize the flap motion and the circular motion of Kinect. The efficiency of the scanning has been improved by automatic control. The optimized scanning parameters have been determined via various scanning tests. Compared to the hand-held scanning, the experimental results indicate that the time cost can be reduced, and the quality of point cloud obtained can be improved significantly.

The integration of reverse engineering and rapid prototyping technology has become a relatively independent research field in the CAD/CAM system. As an important and an advanced supporting technology which digests, absorbs and shortens the product redesign and manufacturing cycle, in recent years it has become the hot spot for manufacturing industry. Taking a hook claw as an example, this paper describes a typical part of the point cloud acquisition, based on the studio Geomagic model reconstruction, and the process of the prototype production via 3D printer. And the key issues such as point cloud processing, polygon processing and surface processing are discussed in the process of model reconstruction, and the error analysis of data is presented and the method of reducing error is presented.

Integrated inspection system (IIS) enriches digital metrology by decreasing uncertainty due to closed-loop implementation of the cyber-components of digital metrology including planning the measurement of points, fitting the best Substitute Geometry (SGE) and evaluating the actual geometry of the surface. Optical devices are favored over contact metrology devices as they offer faster, fuller, and non-invasive measurement, but they are also susceptible to noise and outliers that may affect the accuracy and quality of the inspection. Principal Component Analysis (PCA) has been presenting high efficiency in attaining transformation matrix of the large number of discrete points given by the optical devices based on the point’s distribution; however, it is not robust to noise. Noise removal prior to SGE provides pure information feedbacked to the other cyber-components of IIS, and finally correctly identifies the actual geometry of the inspected surface. In this paper, in order to obtain the robust principal component of the unorganized point cloud of the planar surfaces, a subset of the data, which is least affected by noise, is found with Minimum Covariance Determinant (MCD), and then PCA is conducted on the obtained MCD-subset. Due to the MCD’s robustness, the noises can be detected by their large robust distances based on the MCD location and scatter matrix. The results of implementation on two case studies show the efficiency of MCD in simultaneous noise removal and robust surface fitting. Variety of inspection tools can utilize the methodology for point cloud filtration.