Research On The Detection Pose Estimation And Verification Recognition Technologies Of Point Cloud Faces In Complex Scenarios

In the field of computer vision,head detection,head pose estimation,and face verification and recognition techniques have always been the focus of research.Breakthroughs have been made in related research based on ordinary 2D images,which have been widely applied in daily life and work.However,the imaging quality of ordinary 2D images is easily affected by changes in lighting,making them ineffective for applications in scenes lacking light or with large variations in lighting.With the rapid development of 3D sensors such as lidar and infrared cameras,point clouds have gained increasing attention in the field of computer vision due to their advantages of stability against changes in lighting and rotational invariance.Based on 3D point cloud technology,this dissertation conducts research on head detection in complex backgrounds,head pose estimation under partial occlusion,face verification under pose interference,and face recognition problems,and has made the following progress.In order to solve the problem of head detection in complex scenes,this dissertation proposes a new head detection network Point Head Net by clustering analysis.The network initially divides the point clouds space into a set of equally sized cubes,and treats each cube as a voxel in the 3D space,thus encoding the point clouds space into an ordered mapping images,and then uses a Region proposal network to conduct preliminary screening on the detection targets.Next,the mean shift algorithm is used for clustering analysis of candidate regions,aiming to eliminate false detections caused by cluttered backgrounds,head pose changes,and partial occlusion,and accurately calculate the center of the head.Point Head Net combines the perceptual ability of the region proposal network for target detection with the clustering ability of the mean shift algorithm to achieve head detection in complex point cloud scenes.Experimental results on multiple public datasets verify the effectiveness of Point Head Net,with an intersection over union(Io U)of over 0.90.To address the problem of head pose estimation under partial occlusion,this dissertation uses multiple networks to constrain the pose features and proposes a new head pose estimation network Point Pose Net.The network first constructs a backbone network for pose estimation based on Point Net++,and uses an Ordered regression network to restrict the predicted pose angles to a more precise range,effectively reducing the prediction error.Subsequently,a Siamese network is introduced to constrain samples with similar pose features,improving the network’s prediction ability for partially occluded samples.Experimental results on public data show that the Ordered regression network and Siamese network reduce the average prediction error of Point Pose Net by 9.09% and 2.70% respectively,and the connection between each subnetwork adopts feature and weight sharing,which takes into account the real-time requirements of the task.To address the problem of point cloud face verification under pose interference,this dissertation proposes a new face verification network Point Siamese by scaling features.The network first constructs a feature extraction network using x-conv operator to obtain facial features,and maps the face features to a fixed radius hypersphere surface through scaling parameters to reduce the interference of pose changes on face verification.Then,Point Siamese uses the Chamfer distance between point clouds to construct a new loss function to measure the similarity between features,further improving the accuracy of verification.Experimental results on public datasets show that the feature scaling parameter and the new loss function respectively improve the verification accuracy of Point Siamese by 2.55% and 1.97%,and can effectively deal with the interference of pose changes.To address the problem of face recognition under pose variations,partial occlusions,and expression changes,this dissertation proposes a new face recognition network called Local Conv,which is built upon local feature descriptors and a feature enhancement mechanism.The network first provides a detailed description of the geometric features of the face using feature descriptors.Based on these descriptors,a new Ψ-conv is constructed to extract fine-grained facial features.Subsequently,a novel feature enhancement mechanism is introduced to further enhance the discriminability of facial features.Experimental results on public datasets demonstrate that the local feature descriptors and the feature enhancement mechanism improve the recognition accuracy of Local Conv by 1.92% and 8.00% respectively,while effectively handling interference from head pose variations,partial occlusions,and expression changes.