Research On 3D Scene Object Perception Technology Based On Multi-view

Visual perception technology is an important support in the field of 3D display technology and artificial intelligence.In recent years,the application of visual perception in artificial intelligence technology has become wider and deeper.Compared with other sensor perception technologies,camera perception-based visual perception has the advantages of low cost,high algorithm flexibility,and simpler application.This thesis focuses on the two important areas of image recognition and 3D vision.In the body of this thesis,the visual perception technology of 3D scene will be studied and discussed from these two aspects.In the traditional 3D reconstruction technology,the restoration of the surface texture of the object is more concerned,but the semantic understanding of the scene is lacking.On the other hand,3D information is ignored in traditional image object detection technology.The research in this paper combines the two.Multi-view 3D information is used to assist image detection,and the results of image detection are used to calculate semantic 3D point clouds to complete the perception of real scenes.The main work of this thesis is as follows:1.The existing image target detection technology and semantic segmentation technology are thoroughly studied and discussed.Their advantages and disadvantages are compared and analyzed.The common features and characteristics of different methods are summarized,which provides a theoretical basis for the work of this paper.At the same time,the internal connection between the three-dimensional scene and the plane image is analyzed,and the feasibility of combining the two is theoretically discussed.2.Based on the analysis of the internal relationship between the object image points and the entire object structure,a multi-view feature point matching algorithm is proposed which fuses semantic information and three-dimensional information and combines closed-loop matching.The algorithm has high robustness and reliability,and has the ability to obtain enough correct information for most scenarios.In this study,these multi-view matching features will be used to relocate targets not detected in a single view.3.A 3D point cloud computing method based on multi-view sliding window is proposed.Combining image semantic detection and feature matching technology,3D points of different objects can be correctly perceived and distinguished.This method can not only avoid a large number of redundant calculations caused by processing all views at the same time,but also can improve the scalability of the algorithm,making it suitable for tasks requiring higher accuracy and higher reliability,such as 3D scene reconstruction.