Scene Analysis And Reconstruction Based On Machine Learning

With the continuous development of computer vision technology,people are interested in simulating the real world of three-dimensional scene with computer,at the same time,the three-dimensional scene reconstruction technology requirements are getting higher and higher.The traditional three-dimensional modeling software inefficient,high learning costs shortcomings become an obstacle on popular three-dimensional scene modeling technology.In order to realize the popular and efficient 3D scene modeling technology,this paper presents a new method to reconstruct the indoor 3D scene from single or multiple RGBD images.First of all,this paper refers to an interactive image semantic segmentation and labeling algorithm,segmenting and labeling on the RGBD image to obtain the scene elements.On this basis,this paper presents a three-dimensional reconstruction method based on model matching,which combines the image blocks with the existing three-dimensional model of the indoor scene,and obtains the corresponding three-dimensional elements of the indoor scene,which is applied to the 3D scene reconstruction.This method is adjusted and trained on the basis of popular convolution neural network in recent years,and a very good matching model is obtained.Aiming at the problem of attitude estimation of three-dimensional elements in 3D reconstruction process,this paper presents a depth graph attitude estimation algorithm based on convolution neural network.The algorithm uses a regression network to estimate the attitude.First of all,we have synthesized a large number of different postures of the depth image through the existing 3D model,so as to solve the problem of training data that requires dense sampling.Then,for different categories of objects,linear regression estimates were used to fit the attitude function.The structure of the convolution neural network is modified on the LeNet-5 model.The network model is designed by using the three-layer convolution layer and the three-layer down-sampling layer,which makes the network suitable for the regression estimation.Finally,the optimal training model is obtained,and the problem of attitude estimation in 3D scene reconstruction problem is solved successfully.The experimental results show that our method achieves an estimate with an average error of about 4.3 °,which is superior to other literature.