Research On Intelligent Position Perception Method For VR Visual Discomfort

Virtual Reality(VR),as a brand-new visual display technology,can immerse users in a virtual stereoscopic environment through devices such as Head-Mounted Displays(HMD),bringing new visual experience to users.,which meets people’s requirements for new information presentation methods,has gradually become an important research topic in the fields of computer graphics,computer vision,and motion perception.However,while VR display devices bring people stereoscopic information,it often causes visual discomfort to users,which seriously restricts the development of VR.Therefore,it is of great significance to carry out research on visual discomfort.This paper studies the visual discomfort caused by inaccurate positioning of VR systems,and is committed to building a Visual Discomfort Prediction(VDP)model to predict the score of visual discomfort when users view stereo images,and to evaluate the effect of depth differences that cause inaccurate positioning in virtual space.At the same time,the positioning algorithm of virtual space and real physical space is studied to improve the accuracy of related positioning algorithms,and to provide theoretical support for enhancing visual comfort and realizing friendly human-computer interaction.It includes the following research contents:(1)Research the VDP model based on image features.First,combined with the visual attention mechanism,the scene structure features and depth difference features are designed and extracted,and then a VDP model based on scene structure and depth difference is constructed,which can objectively assess the impact of depth differences on visual discomfort while predicting visual discomfort scores.Then,in order to predict the visual discomfort score more accurately,Two-path Feature Fusion and Multi-task Visual Discomfort Prediction(TFMVDP)model is proposed.Inspired by the feature integration mechanism of the Human Visual System(HVS),the model constructs a dual-path feature fusion structure to mine more visual discomfort-related image information to represent human visual discomfort perception.In addition,the multi-task learning is adopted.The classification task of discomfort level and the regression task of discomfort score are considered at the same time,and the prediction of discomfort score is guided by predicting the category of discomfort level,which further improves the prediction performance of the model.Experiments show that the proposed TFMVDP model has better prediction result than the existing VDP model.(2)Aiming at the problem of inaccurate positioning in virtual space caused by the inconsistency between the user’s perceived depth and the depth set by the VR system,a selfpose estimation method based on HVS depth perception is proposed to accurately locate the user’s visual system perception of self-position in the virtual space.The proposed method uses the gaze tracking algorithm to establish the user’s depth perception model,and obtains the perception depth of the HVS.After that,the perceived depth is used to construct the internal parameter matrix relative to the human eye,which is combined with the camera’s internal parameter matrix to adaptively obtain the set depth of the VR system in the virtual space without marked points to obtain the set depth,so as to obtain the difference between perceived depth and setting depth,and finally obtain the user’s perceived self position correctly according to the depth difference.In addition,when the gaze tracking is lost,combined with the depth information perceived before the gaze tracking is not lost and the image information of the local area of the gaze point,the depth difference corresponding to the local image area is weighted and calculated,so that,the user’s perceived self-position can still be obtained by using the depth difference of the local image area,when the sight tracking is lost.Sufficient subjective visual discomfort experiments show that the improved VR system according to the proposed method can effectively enhance the visual comfort.(3)Aiming at the problem that the Simultaneous Localization and Mapping(SLAM)method is inaccurate in the complex real physical space with repeated or similar textures,a robust pose estimation method using long-distance feature description and semantic block optimization(Long Distance Features Description and Semantic Block Optimization,LDFDSBO)is proposed to accurately locate the self-pose in the reality space.In the feature matching of SLAM,the LDFD-SBO method generates a feature descriptor(Long Distance Features Description,LDFD)with long-distance dependencies by capturing the long-distance pixel information on the cross path of image feature points,which enhances the e distinguishability of feature points in similar or repeated regions,and can reduce the mismatch of features.At the same time,in the pose optimization of visual SLAM,the LDFD-SBO method takes the region blocks of different semantic categories on the image as units,and adopts the information entropy theory to adaptively obtain the correlation between the feature points of different categories of semantic blocks and the pose estimation,and pay more attention to the feature points in the semantic block area with high correlation,the localization failure caused by indiscriminate processing of image feature points is avoided.The experimental results demonstrate the excellent performance of the proposed algorithm.