Research On Anti Noise End-to-end Binocular Stereo Matching In Internet Of Vehicles

The Internet of vehicles is a research hotspot of intelligent transportation.In the Internet of vehicles environment,binocular stereo vision is used to accurately judge the position of objects,so as to provide technical support for effectively solving problems such as traffic safety.At present,most binocular stereo matching algorithms use end-to-end binocular stereo matching algorithm based on convolutional neural network.However,the problems of large image noise and neural network parameters can not meet the requirements of high precision and high running speed of the Internet of vehicles.This thesis carries out research work on these problems.Aiming at the problem that the vision sensor works for a long time and the image noise is caused by poor brightness during shooting,the accuracy of stereo matching convolution neural network is reduced.Based on LIF biological neuron model with anti noise,this thesis proposes NST function,adjusts the anti noise parameters of NST function to fits the response mechanism of LIF model.NST function has anti noise performance and good sparsity,gradient,mean characteristics and non monotonicity.Simulation results show that NST function is stable when there is noise input,and when the variance value is 0.2,the error value of it in GC-Net is 7.085% lower than Re LU function,and it 8.504% lower than that of Re LU function in PSM-Net.It can effectively overcome the impact of image noise on stereo matching and improve the accuracy.Aiming at the problem that the end-to-end binocular stereo matching algorithm based on convolutional neural network mostly uses 3D convolution kernel cost aggregation,which leads to a large number of network parameters and slow running speed,this thesis proposes a Semi-global Tree Structure Cost Aggregation(SGT)algorithm,which uses the spanning tree to aggregate the generation values from multiple directions to replace the 3D convolution kernel to reduce the network parameters and improve the parallax processing speed;Adjust the parameters of cost aggregation in each direction in SGT algorithm,flexibly control the extraction of image geometric texture information,and effectively solve the problem of high false matching rate in weak texture areas;The similarity between pixels is calculated by pixel gradient to improve the accuracy of image edge disparity matching.In this thesis,an End-to-End Semi-global Tree Structure Cost Aggregation Disparity Prediction Neural Network(SGT-Net)is constructed by using SGT algorithm.The features are extracted at three different scales to capture the full-text information more comprehensively;Using epipolar constraint to construct 4D cost volume,SGT-Net has geometric constraint network and obtains more accurate disparity map;SGT cost aggregation layer reduces the volume of neural network and improves the running speed of neural network.Compared with GC-Net,the running speed of SGT-Net with NST activation function is increased by 0.55 s,the accuracy is increased by 0.455%,and the running speed is improved by 0.13 s and the accuracy is improved by 0.101% compared with PSM-Net.Simulation experiments show that SGT-Net can effectively improve the matching accuracy and speed up the running speed for weak texture areas,which is more suitable for the Internet of vehicles environment with noise impact.