Research On Recognition And Registration Methods Based On Deformed Markers

The recognition and registration method based on fiducial markers is one of the main research problems in augmented reality.However,existing detection methods assume that the markers are placed on ideally planar surfaces and the position of markers is somewhat restricted.In real-world,markers are susceptible to the complex backgrounds,uneven illumination,shooting angles,distances,etc.,and which are prone to deformations such as folding,twisting,and creasing,and thus cannot be recognized.In order to solve the above problems,this thesis takes the Ar Uco markers as the research object and proposes a recognition and registration method based on deformed markers,which combines two deep learning networks with the Pn P algorithm to preform pose estimation and registration of the deformed Ar Uco markers in real-world complex scenes.The main research contents are as follows:(1)Aiming at the problem of markers detection in complex scenes,this thesis proposes a method to detect deformed markers.Firstly,the U~2Net network is studied and improved,and a soft attention mechanism module is integrated into the skip connection structure of the U~2Net to improve the segmentation accuracy of the markers in the scene by highlighting the salient features in the skip connection and suppressing the background irrelevant regions through the idea of weight assignment.Then,the markers are detected and the location information of the markers is obtained.It is verified through experiments that the proposed marker detection method can quickly and accurately detect the deformed markers in complex scenes.(2)Aiming at the problem of deformation correction of the detected markers,this thesis constructs a correction network based on improved encoder for deformed marker images.First,the position coordinates of the control points and reference points of the markers are predicted by the network.Then,the sparse mapping between the control points and the reference points is converted into an inverse mapping using TPS interpolation method.Finally,the correction is achieved by remapping the deformed markers image.In order to better retain the local detail information of the images,this thesis constructs the dilated residual block based on the structure of the residual block.In addition,a spatial pyramid structure is constructed in the network.By stacking different scales of dilated convolutions to obtain multi-scale information of the image without increasing the depth of the network.It is verified through experiments that the proposed method can correct different types of deformed marker images while ensuring the integrity of the coding information inside the markers and improving the recognition rate of the markers.(3)Finally,this thesis produces a dataset containing complex deformed marker images.In order to increase the diversity of marker deformation methods,a dataset containing complex deformed markers is designed.The deformed markers are generated by distorting the standard marker images with a 2D grid,and then preform data augmentation to increase the complexity of the dataset.In addition,a real-world scene dataset is collected,in which the Ar Uco markers are placed on arbitrary planes or surfaces in different backgrounds and illuminated scenes.The annotation was performed manually,and the annotated dataset was collated and analyzed.The experimental effect of the registration of virtual objects leads to the conclusion that the recognition and registration method based on deformed markers proposed in this thesis can effectively improve the recognition rate of markers and the accuracy and robustness of 3D registration.