基于改进PV-RCNN++算法的三维点云聚焦式特征研究

Research on 3D point cloud focusing features based on improved PV-RCNN++algorithm

为增强RoI网格局部特征的表征能力、加强细节特征的表达效果,以进一步提高点云语义分割精度,针对PV-RCNN++网络采用的RoI网格池化模块进行了研究。在PV-RCNN++网络中,RoI网格池化模块只是将网格点周围的体素特征进行简单的空间位置排序,导致局部特征表达效果欠佳。为加强RoI网格池化模块对局部特征的表征能力,引入CBAM注意力机制,从通道和空间两个作用域出发,一方面处理特征集通道的分配关系,另一方面可使神经网络更加关注特征集中对分类起决定性作用的体素区域,以强化重要信息在网络的有效传递并提高点云语义分割结果的鲁棒性。对自动驾驶领域公开数据集Kitti的语义分割实验表明,所提出的改进PV-RCNN++的聚焦式特征的算法训练出的模型,较基准模型提升效果显著,有效增强了RoI网格池化模块对局部特征的表征能力,强化了细节特征的表达效果,提高了点云语义分割精度。

In order to enhance the representation ability of local features of RoI grid,enhance the expression effect of detailed features,and further improve the accuracy of cloud semantic segmentation,the RoI grid pooling module used in PV-RCNN++network was studied.In PV-RCNN++network,RoI grid pooling module only performs simple spatial position ordering of voxel features around grid points,resulting in poor local feature expression effect.In order to enhance the representation ability of RoI grid pooling module for local features,CBAM attention mechanism was introduced.The CBAM attention mechanism starts from two domains of channel and space.On the one hand,it deals with the distribution relationship of feature set channels;on the other hand,it enables the neural network to pay more attention to the voxel region where feature set plays a decisive role in classification,so as to enhance the effective transmission of important information in the network and improve the robustness of the semantic segmentation results of the high point cloud.Semantic segmentation experiments on Kitti,an open data set in the field of automatic driving,show that the model trained by the proposed improved PV-RCNN++focusing feature algorithm has a significant improvement effect compared with the benchmark model,which effectively enhances the representation ability of RoI grid pool module for local features,strengthens the expression effect of detailed features,and improves the semantic segmentation accuracy of point cloud.