Improved point-voxel region convolutional neural network for small object detection
Improved point-voxel region convolutional neural network for small object detection
复制标题
DOI：
10.1117/12.2657106
复制
发表时间：
2022-12
期刊：
影响因子：
--
通讯作者：
Zherui Xie;Masaaki Tsuzaki;Huimin Lu;S. Serikawa
中科院分区：
文献类型：
--
作者：
Zherui Xie;Masaaki Tsuzaki;Huimin Lu;S. Serikawa
随着LiDAR传感器的广泛使用，通过3D点云数据处理进行3D物体检测已成为机器人和自动驾驶领域的研究目标。然而，点云数据的无序性和稀疏性是传统点云数据处理中存在的问题。使用大量的点云数据来检测对象是具有挑战性的。传统的3D物体检测器主要有基于网格的方法和基于点的方法。PV-RCNN提出了一个结合基于体素和基于点的技术的框架，并使用3D体素CNN提取对象特征。然而，CNN引起的分辨率降低会影响对象的定位。这项研究旨在通过不仅向RPN提供体素CNN的单个输出，还向RPN提供包括高分辨率输出在内的多个输出，来提高更多次要事物的检测精度。我们提出了一个新的网络，它引入了多尺度区域建议网络，以减少分辨率下降的影响。我们的网络对自行车等小物体的识别准确率比原始的PV-RCNN更高。在广泛的实验中，我们证明了我们的模型在小事情上有5%的改进，例如在KITTI数据集上训练骑自行车的人。
With the widespread use of LiDAR sensors, 3D object detection through 3D point cloud data processing has become a research target in robotics and autonomous driving. However, the disorder and sparsity of point cloud data are the problems in traditional point cloud data processing. It is challenging to detect objects using a large amount of point cloud data. Conventional 3D object detectors have mainly grid-based methods and point-based methods. PV-RCNN proposed a framework that combines voxel-based and point-based techniques, and object features are extracted using 3D voxel CNNs. However, the resolution reduction caused by the CNN affects the localization of objects. This study aims to improve the detection accuracy of more minor things by feeding not only a single output of the voxel CNN but also multiple outputs, including high-resolution outputs, to the RPN. We came out with a new network that introduces the Multi-Scale Region Proposal Network to reduce the effect of resolution degradation. Our network has better recognition accuracy for small objects like bicycles than the original PV-RCNN. In extensive experiments, we demonstrate that our model has a 5% improvement for small things, such as cyclists training on the KITTI dataset.