Research On Binocular Stereo Matching Algorithm Based On Siamese Convolutional Neural Network

Binocular stereo vision technology simulates the human visual system’s spatial perception of the real world to construct the structure of objects in three-dimensional space.At present,this technology is widely used in unmanned aerial vehicle reconnaissance,robot control and navigation,automatic driving,three-dimensional reconstruction,and object recognition and detection.Stereo matching is an extremely critical step in binocular stereo vision,and its accuracy will have a huge impact on the effect of 3D reconstruction.Therefore,improving the accuracy of the stereo matching algorithm in the binocular stereo vision system is of great significance to the development of computer vision in the future.This thesis mainly studies the binocular stereo matching algorithm,and integrates the convolutional neural network into the matching cost calculation step of the stereo matching algorithm.A stereo matching algorithm based on Siamese convolutional neural network is proposed.The main research contents are as follows:(1)A convolutional neural network structure based on Siamese model is proposed to calculate the matching cost.The algorithm is divided into two parts: feature extraction and feature calculation based on the Siamese model.It uses the shared weights on both sides to perform multi-layer convolution,and feeds the output feature planes to the hollow convolutional layer and the spatial pyramid pooling layer,and then extracts image features by cascade and multi-layer convolution,and finally initializes the matching cost by calculating the cosine similarity.In traditional classic algorithms,the matching cost often requires complicated manual design to extract features.In contrast,the method in this thesis is obviously more concise,and can get more accurate matching costs.(2)The cost aggregation algorithm based on crossover and semi-global is used to aggregate the matching cost of the previous step.In the aggregation algorithm based on the fixed window,the weighted average of the single pixel matching cost in each fixed window need to be calculated,but this method has great limitations in the discontinuous area of the disparity.The cost aggregation algorithm adopted in this thesis effectively reduces the mismatch rate in discontinuous areas of disparity depth by increasing the conditions for constructing cross arms,expanding the cross coverage area and specifying different aggregation directions during iteration.Then,using the Winner-takes-all strategy to calculate the disparity.Finally,post-processing methods such as Left-Right Check,Subpixel Enhancement,and Median Filter are used to optimize the disparity map.Experimental results show that the algorithm in this thesis can generate more refined disparity maps on the KITTI and Middlebury test platforms.