A Stereo Matching Method Based on Deep Learning

Abstract

The present invention provides a stereo matching method based on deep learning, using the deep learning method to extract features, the receptive field of each point of the obtained feature spectrum is very large, but at the same time there is no foreground expansion effect brought by the traditional local matching method, Solved the problem that the partial matching method window was difficult to select. Moreover, the high-dimensional feature spectrum obtained through deep learning fully describes the data distribution characteristics of the region, which has a strong guiding significance for the subsequent similarity detection. According to the left and right views, the present invention uses a deep learning method to output a W*H*D cost body, and then uses the winner-takes-all WTA strategy to obtain an initial disparity map, and finally uses the original left view as a guide image for guide filtering. This method has a higher Accuracy, and due to the high degree of parallelization of the neural network, the reasonable use of GPU parallel acceleration can make the method achieve a very fast computing speed.

Description

technical field [0001] The invention relates to stereo matching technology. Background technique [0002] Binocular vision is an important branch of computer vision, and stereo matching is the most important part of binocular vision. Binocular distance measurement generally uses two identical cameras to obtain two images of the same scene with different perspectives at the same time, and then uses the stereo matching method to find the corresponding relationship between the same object or texture in the two images, and finally restores the image based on the parallax. The distance of the object from the camera. [0003] In general, stereo matching methods can be divided into global methods and local methods. Based on the assumption of smoothness, the global method uses various methods to construct the global energy function, and uses the optimization method to calculate the minimum point of the energy function. Whereas local methods rely on local information and determine...

AI Technical Summary

Problems solved by technology

[0004] The global stereo matching algorithm uses two constraints of image smoothness and data items to solve the minimum value of the overall energy function, which usually has a large amount of calculation, and is difficult to perform parallel optimization, takes a long time, and is not suitable for real-time systems

[0005] Semi-global stereo matching has the following defects: (1) It is difficult to determine the size of the window: if the window is too large, it will require a large amount of calculation, and the foreground expansion effect will be generated in the depth discontinuity area, and the effect of the edge is very poor; if the window is too small, it cannot contain All the texture features of the object cause ambiguity in the matching results

(2) The value used to calculate the cost is only the gray value of the image or the low-order difference, which cannot describe the features well

Even if the non-parametric Census transform is used, it only enhances the robustness to noise and cannot solve the problem of insufficient feature description.

Images