Research On Depth Map Acquisition Based On Binocular Vision

With the advent of the era of artificial intelligence,deep map acquisition technology has become a hot research direction in the field of computer vision,and it is the core technology for the perception of 3d scene by machine equipment.The key of binocular depth map acquisition technology is stereo matching.In order to obtain more accurate depth map in less time,this article focuses on efficient and accurate stereo matching algorithms.The specific research contents are:(1)The edge graph generated by the previous edge detection algorithm based on deep learning is rough and fuzzy in details,and there are errors in the constraints of stereo matching.To solve these problems,this paper proposes an efficient edge detection algorithm based on convolutional neural network as the edge constraint of stereo matching algorithm.In order to improve the model of feature extraction ability,this algorithm introduce the mechanism of "attention" into the backbone network and use the Squeeze-and-Excitation module to extract the image edge features.In order to retain more detailed information of the edge image,this algorithm removes part of the down sampling in the backbone network to improve the feature resolution,and uses the extended convolution technology to increase the sensing field of the model.In order to fully fuse the multi-scale features of the edge,this algorithm designs a feature fusion module based on residual structure,fuses the edge graphs of different scales to improves the precision of the edge graphs.The experimental results show that the edge detection algorithm can output finer edges and is more robust to noise such as illumination and shadow.(2)In order to solve the problems of excessive smoothness and fuzziness in the edge region of depth map,this paper proposes an efficient stereo matching algorithm based on edge constraint.This algorithm integrates the stereo matching network based on the spatial pooling pyramid with the high-precision edge detection network to form a multi-task learning mechanism.In order to extract more geometric and detailed features from the stereo matching branch,an efficient feature fusion module is designed to integrate the multi-scale edge features from the edge detection branch into the stereo matching branch.In order to make the stereo matching branch correctly perceive the edge region and non-edge region in the image,the algorithm takes the edge graph generated by the edge detection branch as the edge guidance and adds a penalty term of edge perception in the loss function of the stereo matching branch to punish the behavior of wrong edge perception.Experimental results show that the stereo matching algorithm can greatly alleviate the problems of excessive smoothing of edge region,small target and dense region in depth map.(3)Aiming at the large number of parameters and poor portability of the stereo matching algorithm based on neural network,this paper proposes a lightweight fast stereo matching algorithm based on edge constraint.In order to solve the problem of parameter redundancy in feature extraction network,this algorithm proposes a lightweight feature extraction network based on Shuffle V2 module,which reduces the number of parameters of the module and improves its expression ability.In order to reduce the number of 3D convolution operators in the similarity computing network,this algorithm improves the calculation method of similarity and uses 2D convolution to replace 3D convolution,decouples the correlation of disparity dimension,and improves the generalization ability of the model while reducing the number of parameters.In order to improve the disparity precision of the model prediction is not affected,this algorithm proposes a disparity refinement module based on the attention mechanism to correct the error of the initial disparity and further improve the output of the model.Experimental results show that the algorithm reduces the number of model parameters by half without affecting the disparity accuracy,and improves the portability of the model.