Research On Three-Dimensional Scene Reconstruction Based On Depth Data And Image Semantic Segmentation

In recent years,the study of three-dimensional scene reconstruction has been paid more and more attentions,which makes it possible for the technologies such as three-dimensional digital maps to obtain a vigorous development.Three-dimensional scene reconstruction is a very important research field in computer vision research.With the popularity of depth cameras,three-dimensional reconstruction based on depth data has made very large progress that can help locate the position in a scene and eventually get a complete three-dimensional model.With the rapid iterative development of deep learning,the image semantic segmentation based on convolution neural network has achieved very significant results.Image semantic segmentation can be used to segment the image at the pixel level.Therefore,it can be applied to three-dimensional scene reconstruction based on depth data.and a three-dimensional model with simple semantic segmentation can be obtained.In this thesis,image semantic segmentation based on convolution neural network is applied to three-dimensional scene reconstruction based on depth data,and the related key technologies are studied deeply.In the module of three-dimensional reconstruction based on the depth data,a pose optimization strategy based on sparse matching and dense matching is proposed.Based on the optimization of geometrical and photometrical consistency,the sparse feature optimization term is combined and experimented on the standard data set to verify the effect of the optimization.In the module of image semantic segmentation based on the convolution neural network,the pairwise potential function of the conditional random field is optimized,and a large number of experiments have shown that the segmentation effect has been improved.On the basis of the above technologies,a three-dimensional scene reconstruction solution based on depth data and image semantic segmentation is proposed.The three-dimensional model is obtained by depth data,the semantic segmentation of image pixels is obtained by image semantic segmentation,and the predicted semantic classification labels are migrated to the three-dimensional model by Bayesian based incremental transfer strategy,and finally the three-dimensional model with simple semantic segmentation is generated.In order to obtain a three-dimensional model with simple semantic segmentation,this thesis adopts a three-dimensional scene reconstruction solution combining image semantic segmentation and depth data,but there are still many shortcomings,still need to continue to improve.At the end of this thesis,we have also made a prospect for the improvement and development of the follow-up.