点云投影结合轻量化卷积神经网络实现三维成像声呐快速目标分类

Point cloud projection combined with lightweight convolutional neural network realizes fast target classification of 3D imaging sonar

三维成像声呐的成像结果是三维点云,基于点云的三维成像声呐目标分类方法具有网络结构复杂、计算量大的特点。针对这一问题提出了一种将三维成像声呐成像结果从三维点云投影至二维图像的方法,并且使用轻量化卷积神经网络实现了三维成像声呐快速目标分类。该方法首先对三维成像声呐波束形成后的波束域数据进行最大值滤波和阈值滤波,以降低点云数据维度;接着,依据三维成像声呐的波束方向,将点云投影为深度图和强度图,分别保存点云的位置信息和强度信息;然后,利用深度图和强度图分别作为第一个通道和第二个通道构建混合通道图,将混合通道图作为目标分类网络的输入,从而将三维点云的目标分类问题转换为二维图像的目标分类问题;最后使用MobileNetV2网络实现了三维成像声呐快速目标分类。实验结果表明,提出的投影方法可以用二维图像分类网络完成三维成像声呐点云的目标分类任务;而且训练过程中混合通道图比单独的深度图和强度图收敛速度更快,结合目标分类网络可以实时进行目标分类,在真实数据集上分类正确率达到了91.13%。

The imaging result of the three-dimensional(3D)imaging sonar is a point cloud,and the network of point cloud target recognition is characterized by complex network structure and large computation.We propose a method to project the imaging result of three-dimensional imaging sonar from a point cloud to an image,and use lightweight convolutional neural networks to achieve fast target classification for three-dimensional imaging sonar.Firstly,the method performs maximumfiltering and thresholdfiltering on the beam domain data after the beamforming of the 3D imaging sonar beam to reduce the dimensionality of the point cloud.Next,based on the beam direction of the 3D imaging sonar,the point cloud is projected to a depth image and an intensity image to save the point cloud position information and intensity information respectively.Then,the mixed image is constructed using the depth image and the intensity image as thefirst channel and the second channel,and the mixed image is used as the input of the target classification network,thus converting the target classification problem of 3D point clouds into the target classification problem of images.Finally,3D imaging sonar fast target classification was implemented using MobileNetV2.The experimental results show that the projection method proposed in this paper can be used to complete the target classification task of three-dimensional imaging sonar point cloud by an image classification network.Moreover,the convergence rate of the mixed channel image is significantly faster than that of the separate intensity image and depth image,and the target classification can be conducted in real time with the combination of the target recognition network,achieving an accuracy of 91.13%on the real data set.