Research On Recognition And Detection Of 3D Point Cloud Based On Deep Learning

The recognition and detection of 3D point cloud data is a research hotspot in the field of computer vision.It has important applications in many important fields,such as unmanned driving,high-precision maps,and robot-assisted vision.At the same time,with the research and development of deep learning technology,the research of 3D point cloud data recognition detection combined with deep learning technology has attracted more and more attention and attention.The Pointnet network pioneered the use of deep learning technology to directly process 3D point cloud data,and realized the classification and segmentation of 3D point cloud models.However,the disadvantage is that the feature information between points can not be extracted well,which makes the ability of Pointnet network to capture local features is not strong;at the same time,in the application of Pointnet,frustum Pointnet network model can achieve the detection of three-dimensional point cloud targets,but due to the low efficiency of Pointnet network to point cloud segmentation,frustum Pointnet detection accuracy is still greatly improved Space.In this paper,three-dimensional point cloud recognition and detection based on deep learning is studied,and the main work is as follows:(1)In order to improve the Pointnet network,the region features are extracted by dividing the input 3D point cloud data into regions and replacing the single point of the Pointnet network with the point cloud of the region.Aiming at the requirements of completeness,adaptability and overlap of local area division,an area division algorithm based on FPS and KNN is proposed.At the same time,the multi-scale method is used,the K value of the KNN algorithm is used as the scale parameter,and the point cloud regions of different scales are obtained by setting different K values,and then the single-scale point cloud region feature extraction network is constructed.Fusion of feature combinations of different scales to extract multi-scale fusion features.Finally,through the multi-layer perceptron and tandem pooling operation,the global features are extracted and sent to the classifier to complete the classification andsegmentation tasks of the 3D point cloud model.In this paper,the improved model is named KT-Pointnet network.Experimental results show that compared with the PointNet model,this model exceeds the original Pointnet network model in the classification task on the Model Net40 dataset and the segmentation task in the Shape Net dataset.(2)In view of the shortcomings of point cloud data in frustum Pointnet,which is a 3D object detection application model of Pointnet.The first point of improvement is to embed a new feature extraction operation EdgeConv in Pointnet point cloud segmentation network,calculate the edge features between points and neighboring points and the features of each independent point in point cloud data to capture points The second point is to embed edgeconv in Pointnet modeless 3D frame evaluation network in frustum Pointnet.By embedding EdgeConv,the shortcomings of Pointnet based point cloud segmentation network and Pointnet based modeless 3D frame evaluation network in Frustum-Pointnet network are solved.The experimental results show that the accuracy of frustem dgnet is more than 2% higher than that of the original frustum pointnet method under the condition of difficult samples.