基于动态图特征的堆叠宽度学习三维物体识别网络

Stacking Broad learning 3D object recognition network based on dynamic graph features

三维物体点云识别是智能机器人环境感知任务中的重要组成部分。提出一种基于动态图特征的堆叠宽度学习三维物体识别网络(DG-S-BLS),利用动态图卷积网络提取点云的高维特征,通过宽度学习系统(BLS)模型依据样本整体特征对点云分类,再通过基于BLS块间残差的堆叠宽度学习系统模型进一步提高分类精度。在Li DARNet户外点云数据集上的实验结果表明,DG-S-BLS的分类准确率可达99.5%。

3D object point cloud recognition is an important component of environment perception tasks for intelli-gent robots.Based on dynamic graph features,this paper proposes a Dynamic Graph Stacked Broad Learning System(DG-S-BLS)network for 3D object recognition.DG-S-BLS extracts high-dimensional features from point clouds u-sing a dynamic graph convolutional network,and then uses the Broad Learning System(BLS)model to classify point clouds based on the overall features of samples.The classification accuracy is further improved by using the Stacked BLS model performed upon the residual of the BLS blocks.Experimental results on the LiDARNet outdoor point cloud dataset show that the classification accuracy of DG-S-BLS reaches 99.5%.