Point Cloud 3D Object Detection Technical Research Based On Deep Learning

Three-dimensional(3D)object detection is a fundamental and core task in the field of 3D vision,and it is also a key technology for 3D scene understanding.It has a broad application prospect in the fields of biomedicine,robot navigation,military field,remote sensing mapping,augmented reality and so on.Compared with twodimensional(2D)images,3D point cloud has the advantages of containing the geometric structure and spatial position information of objects and cannot be affected by the change of light texture easily.However,due to the sparse and irregular characteristics of point clouds,the deep learning method based on traditional convolutional networks cannot directly be applied on point clouds.In addition,the complexity of indoor scenes and occlusion problems,increase the difficult of 3D object detection.In view of this,this paper carries out research on indoor point cloud 3D object detection task,and the specific research content is as follows:Firstly,aiming at the problem that the point cloud features extracted by the existing research methods are not refined enough,a refined point cloud feature extraction method based on point-cluster is proposed.The point-based refined feature extraction module is used to further capture the spatial point relationship and shape information,and the refined point cloud deep features are extracted.At the same time,local features are constructed using local grouping strategy to form point clusters.Then the clusterbased feature enhancement module is used to extract the interrelationships between clusters in the high-dimensional feature space of point cloud.The feature enhancement module introduces and the attention layer of multi-layer Transformer is used to enhance the semantic feature information of point cloud to obtain the refined semantic feature information.The proposed feature extraction algorithm was tested on Model Net40 and Shape Net Parts datasets,and the accuracy reached 93.5% and 85.7% respectively,which improved the ability of point cloud feature extraction.Secondly,aiming at the problem of uneven density of point cloud on object surface and lack of feature information,a 3D object detection network based on density information-local feature fusion is proposed.Based on the refined feature extraction method based on point-cluster proposed in this paper,the backbone feature extraction network of 3D object detection is designed,and the density of each point is calculated in the feature extraction stage.The density matrix is learned by using multi-layer perceptron,and the density information is integrated into the feature information of each point to adjust the sampling process of each layer.In addition,attention mechanism is used to encode between local features with density information,so that the network can obtain more abundant feature information of the target object,and effectively improve the performance of 3D object detection.The proposed 3D object detection method is verified by experiments on SUN RGB-D and Scan Net V2 datasets,and the accuracy of 3D object detection reaches 54.7% and 57% respectively,which improves the accuracy of 3D object detection.Finally,for indoor scenes,a 3D data acquisition platform based on mobile robots and solid state lidar is built.The point cloud data of laboratory,conference room,medical treatment and library were obtained through the data acquisition platform,and input into the 3D object detection network proposed in this paper for detection.In the indoor scene with an area of 5m×7m,the detection accuracy can reach 91.7%,which verifies the effectiveness of the algorithm.