Research On Scene Depth And Structure Prediction Method

Deep prediction and structural reconstruction are a hot research topic in the field of computer vision and robotics.The construction of depth maps is an important prerequisite for three-dimensional reconstruction.Traditional methods mainly rely on determining the depth of fixed points for manual annotation or binocular positioning according to the change of camera position to predict depth,but such methods are time-consuming and labor-intensive on the one hand,and it is also limited by factors such as camera position,positioning method,probability of distribution,and other factors.It is difficult to guarantee the accuracy,which makes it difficult to complete the subsequent 3D reconstruction of the predicted depth map.This paper solves this problem by introducing a multi-task module-based deep learning method This problem specifically solves how to achieve the depth prediction of monocular images,combined with semantic segmentation,surface vector prediction,and three tasks to train together,enhance the feature sharing type,and improve the accuracy and structure of depth prediction.This paper proposes a multi-task model-based monocular image depth prediction network for scene images,which can simultaneously train three tasks of learning deep prediction,semantic segmentation,and surface vector estimation,including a common feature extraction module and a cross-task feature fusion module.Extract common features while ensuring the independence of each feature,improving the structure of each task while ensuring the accuracy of prediction.In addition,for the reconstruction of 3D structures,in order to solve the structural redundancy caused by multi-views,to prove that perspective prediction plays a role in promoting the ability of three-dimensional feature extraction and reconstruction quality,this paper proposes a point cloud reconstruction network based on perspective planning.This method based on deep learning,a three-dimensional point cloud reconstruction network for multi-view images is designed.The network includes a 3D geometric semantic feature extraction module,a three-dimensional point cloud reconstruction module and a perspective planning module.The entire network is a unified training architecture.The algorithm is used It can predict the three-dimensional point cloud structure of the image well when the angle of view information is small,and the prediction planning of the angle of view can also increase the efficiency,so as not to cause the problem of excessive calculation.