Key Techniques Of Realistic Details And Interaction In 3D Reconstruction

Three-dimensional reconstruction is a hotspot in several fields such as computer vision,computer graphic and other fields.Recently,the appearance of consumed depth camera(Kinect)makes related techniques of three-dimensional reconstruction gained more attention and re-quired higher accuracy.The 3D model which reconstructed from depth data always accompa-nies low accuracy,lack of realistic details and less intuitive interaction.Therefore,effective methods to reconstruct 3D model with realistic details from captured data and to provide intu-itive interaction approach have significant value of practical application.Due to the fact that frames will be interrupted while the consumed camera is moving,the re-constructed result is not accurate.To solve this problem,this paper proposes a three-dimensional registration method which based on superpixel segmentation.Superpixel segmentation is used to divided global registration into varieties of local registration,which is capable of dealing with the problem caused by the fast movement of cameras and increasing the accuracy of registration.Thereafter,further research shows that the 3D model generated from three-dimensional reconstruction will bring a problem that 3D models have insufficient surface details and the reconstruction result is lack of realistic details.To solve this problem,this paper proposes a method,which estimates the illumination model on key frames,to refine models based on the result of three-dimensional reconstruction.Then,the normal and the position on each point are refined based on estimated illumination model,and the enhancement of realistic details on models is achieved by fusing refined points to original 3D model.Existing interaction methods on reconstructed models are not intuitive and this paper pro-poses a hand gesture interaction method on 3D models to deal with the poor performance of hand gesture recognition on challenging cases such as limited illumination,various skin colors.This paper uses random forest skin detection to segment hand region,and convex hull detection is used to extract hand features and to increase the accuracy of gesture recognition.Eventually,this method is applied in an application where the 3D model is controlled by hand gesture and the performance of this application shows this method's effectiveness.To sum up,this paper analyzes key techniques of realistic details and interaction in three-dimensional reconstruction.Quantitative experiment results show the effectiveness of our method.The methods and achievement in this paper play a catalytic role in further development on three-dimensional reconstruction and related techniques,and they will also facilitate further research on human-computer interaction.