Abstract
Recognizing and predicting future three-dimensional (3D) scenes are crucial steps for real-time vision-based control systems, as these steps enable them to react appropriately in advance. In this study, a method for predicting the position of a 3D point cloud in the future and simultaneously segmenting the predicted point cloud is proposed for the first time. The prediction and segmentation tasks are performed by a novel neural network architecture that extracts both local geometric features and flow features for joint segmentation and prediction. Furthermore, we propose a new evaluation metric for future point cloud segmentation to resolve the problem of inconsistency in the order of future point clouds. The results of experiments conducted using real-world large-scale benchmark datasets revealed that the proposed network achieves higher prediction and segmentation accuracy than other baseline methods.
Highlights
As three-dimensional (3D) vision data can provide abundant spatial information, it is being widely used in many areas, including autonomous driving and mobile robots [1]
The PointRNN+RandLA-Net method that directly input the predicted coordinates into the segmentation network exhibits a poor performance. This is because the prediction network produces errors, which reduce the performance of the segmentation network
Because the segmentation structure of the proposed network is similar to the PointRNN structure, the recurrent neural network (RNN) unit with the geometric feature constraint will reduce the network prediction performance
Summary
As three-dimensional (3D) vision data can provide abundant spatial information, it is being widely used in many areas, including autonomous driving and mobile robots [1]. With recent advances in LiDAR technology that enable quick collection of 3D point clouds with high density and accuracy [2], recognition of point clouds has become a significant research topic. Direct processing of point clouds has been a big challenge because the point cloud acquired by LiDAR is irregularly sampled, unstructured, and unordered in general. Since the introduction of PointNet [3], which was designed to directly extract unordered point cloud features by a symmetric structure, many researchers have proposed PointNet-based deep neural network models that directly utilize point clouds for various tasks including classification, segmentation, and detection.
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
