| In recent years,artificial intelligence has developed rapidly and used in all walks of life.In applications such as mobile robots,augmented reality,and drones,smart devices need to sense their position in the real scene and the three-dimensional structure of the scene.In these applications,high-precision 3D maps can serve their own positioning and prior knowledge.Therefore,how to construct high-precision three-dimensional maps has become one of the hotspots of current research.The three-dimensional map construction scheme based on image sequence only needs to collect data through an ordinary camera.This solution has attracted much attention due to its obvious advantages in terms of price,weight and power consumption.The current three-dimensional map scheme based on image sequence is mainly divided into real-time visual-based SLAM(Simultaneous Localization and Mapping)and SFM(Structure from Motion).However,as the application scenarios become more complex and diverse,existing 3D map construction based on SFM and visual SLAM technology have some problems in the face of complex large-scale scenes.On the one hand,current 3D map construction methods are difficult to estimate the pose of the camera in sparse areas of image texture features.There is a problem of efficiency and precision when optimizing the camera pose,and the drift is serious when constructing a three-dimensional map of a large-scale scene.On the other hand,most existing 3D map construction methods are based on the scanning of handheld devices to obtain sequence pictures.The data obtained by the user's improper operation and other factors cause the data to be discontinuous.These issues will all affect the quality of the constructed 3D map.In view of the above problems,this paper deeply studies the SFM and visual SLAM techniques in indoor complex large-scale scenes,and proposes a series of high-precision three-dimensional map construction methods for indoor complex large-scale scenes for image sequences.Not only is it more robust than the previous method and achieves higher precision camera pose,but it can also obtain high-precision 3D maps.Specifically,the main contributions of this article are as follows:(1)In order to solve the problems that the current three-dimensional map construction method is difficult to have good feature matching in the texture sparse region and it is difficult to estimate the camera pose,this paper proposes a feature matching three-dimensional map construction method using GMS.In the GMS-based feature matching method,Brute Force is used to obtain feature matching pairs,and then statistical matching is used to eliminate the false matching in the number of correct matches in the neighborhood.Finally,the accuracy of the camera pose estimated by the method of this paper and the ORB-SLAM2 method is compared on the TUM dataset.It is proved that this method can obtain high-precision camera poses in large-scale texture sparse scenes.(2)For solving the problems that the three-dimensional map drift constructed by the current three-dimensional map construction method is serious and the range of RGB-D camera measurement is small,this paper proposes a three-dimensional map construction method based on handheld RGB-D camera and surfel model.The method uses GMS feature matching 3D map construction method to obtain higher precision camera pose,2D and 3D feature points as the initial value of the proposed bundle adjustment by combining of 2D and 3D feature points,which can accelerate the convergence of the algorithm.Moreover,the bundle adjustment algorithm uses short-distance high-precision 3D feature matching points as absolute constraints,provides absolute scale information,and combines remote 2D feature point optimization.This can increase the mapping capability of the RGB-D camera.In the end,higher precision camera pose is obtained.Then,the surfel model and the constructed deformation maps are combined and optimized to obtain a high-precision dense three-dimensional map for indoor large-scale scenes.Finally,the experimental results show that the bundle adjustment by combining of 2D and 3D feature points algorithm proposed in this paper can obtain higher precision 3D maps.(3)To solve the problem that the current three-dimensional construction method is mostly based on the handheld camera scanning acquisition,which has high technical requirements for the scanning personnel and large amount of human-computer interaction required,this paper proposes a map construction method based on coordinated scanning of multiple RGB-D cameras.The method uses a fully automated image acquisition platform to ensure the quality of the acquired image.And the advanced image acquisition platform can obtain a higher precision initial value of the positional relationship between the images in advance.For obtaining high-precision 3D maps in complex large-scale scenes,this paper proposes a segmentation cluster adjustment algorithm suitable for multi-site 3D map construction.The algorithm uses each site as a starting segment.Each segment is given a set of motion variables,which are solved independently and optimized,so it is not easy to fall into the local optimal solution.Finally,high-precision camera poses are obtained to create high-precision 3D maps of large-scale scenes.The segmented bundle adjustment algorithm verifies that it can obtain higher precision camera poses through the TUM data set.The accuracy of the three-dimensional map constructed by collaborative scanning of multiple RGB-D cameras is also verified to meet the needs of smart devices and the like.The three-dimensional map construction method proposed in this paper can construct high-precision three-dimensional maps in complex large-scale scenes.The 3D map construction method based on the handheld RGB-D camera is convenient and fast to construct.The three-dimensional map construction method based on rotating rotary platform has low human-computer interaction.The obtained three-dimensional map can be widely applied to fields such as augmented reality and robots. |
PDF Full Text Request