点云目标检测残差投票网络

高精度的三维目标检测是实现物体感知的关键技术, 对自动驾驶、机器人控制等应用的落地具有重要意义。为提高三维目标检测的精度, 对算法VoteNet改进, 提出了一种基于残差网络的端到端的高精度三维点云目标检测网络ResVoteNet。具体来说, 设计了适用于点云数据的残差网络骨架, 提出了残差特征提取模块以及残差上采样模块, 并集成进VoteNet框架。残差网络结构的引入增强了网络对点云数据的特征提取和学习能力, 并且提高了模型的鲁棒性。该算法在公开的大规模点云数据集SCANNET和SUN-RGBD上进行实验, 平均检测精度mAP分别达到61.1%和59.9%, 超越了当前最先进水平的其他算法。

A Residual Neural Network with Voting for 3D Object Detection in Point Clouds

High-precision 3D object detection is a key technology to realize object perception, which is of great significance to the implementation of applications such as automatic driving and robot control. In order to improve the accuracy of 3D object detection, the algorithm VoteNet is improved, and an end-to-end high-precision 3D point cloud target detection network based on residual network, ResVoteNet is proposed. Specifically, a residual network skeleton suitable for point cloud data is designed, and a residual feature extraction module and a residual up-sampling module are proposed and integrated into the VoteNet framework. The introduction of the residual network structure enhances the network's feature extraction and learning capabilities for point cloud data, and improves the robustness of the model. The algorithm is tested on the publicly available large-scale point cloud data sets SCANNET and SUN-RGBD, and the average detection accuracy mAP has reached 61.1% and 59.9%, respectively, surpassing other current state-of-the-art algorithms.