A Research Of Indoor Scene Perception And Reconstruction Based On A Single Panorama

In the field of computer vision,the layout restoration of indoor scenes has always been a core research topic.Early researches on indoor scene layout restoration mostly relied on geometric calculations,however,with the improvement of hardware technology and the development of deep learning methods in recent years,deep neural network algorithms have gradually been used in solving image processing problems.So far,the research on the reconstruction method of indoor scene layout in 2D images has made significant progresses,but the research on the restoration method of indoor scene layout in panoramic images is still very limited.On one hand,because the panorama has nonlinear characteristics,most objects in the panoramic image have various distortions,making it difficult to apply traditional image processing methods directly;on the other hand,because of the omnidirectional nature of the panorama,it is also inevitable to consider the continuous relationship between different orientations when processing panoramic images.However,due to these characteristics,using panorama to restore the layout of the indoor scene can truly show the complete layout and structure of the room,which has great advantages in various applications.Therefore,paper focus on the study of layout reconstruction of indoor scenes based on panoramic images.A method of indoor scene perception and reconstruction based on a single panorama was proposed,and this method based on the existing panorama processing and 3D reconstruction methods to complete the 3D reconstruction,including the frame model and the object model only by single 360 ° panorama of the indoor scene.In this paper,our research methods can be briefly summarized as the following parts.First,the feature extraction algorithm is used to process the panorama of the indoor scene to obtain various structural features of the entire room.Then,using these structural features to complete the frame model reconstruction of the scene.Next,the target recognition algorithm is used to process the panorama to obtain information such as the location and category of all the main targets in the scene,and retrieving the appropriate object model from the model library by comparing the model.Finally,integrating the frame model and object model obtained and completing the reconstruction of the entire indoor scene.In this paper,our research work mainly has the following contributions.The first is to propose an object recognition algorithm for panoramic images.The proposed algorithm overcomes the problem of irregular distortion of panoramic images by adding a panorama processing step to the ordinary object recognition algorithm.The second is to propose a method for resolving the problem of repetitive or false detection of objects in the panorama.The proposed method avoids this problem by introducing a negative feedback mechanism and adopting a method of continuously approaching the correct target object.The third is to propose a model placement method during scene reconstruction.This method combines geometric derivation and model rendering to calculate a reasonable placement of the 3D object model.Compared with the previous indoor scene reconstruction methods,the proposed method not only solves the problem that the reconstructed scene is usually incomplete when reconstructed by perspective images,but also compensate for shortcomings that it is inability to interact with objects in the scene in the current panoramic reconstruction algorithm.At the same time,by adjusting the training data set,the method studied in this paper can also be extended to many other 3D scene reconstructions scenarios.