Deep learning stereo matching method integrating SIFT features

In recent years, convolutional neural networks have developed rapidly, and stereo matching methods based on deep learning have emerged accordingly. Compared with stereo matching methods based on artificially designed criteria, deep -learning-based stereo matching methods can capture details and semantic information in images well without the need to manually design complex feature extraction algorithms. The learning from a large number of samples enables deeplearning- based stereo matching methods to automatically adapt to changes in conditions such as lighting, texture, and viewing angle, and they also have certain robustness in matching ill-posed regions such as regions with repeated lighting textures and weak textures. However, this requires a large number of samples and extreme feature enhancement. At the same time, SIFT features can maintain invariance to image rotation, scale scaling, and brightness changes without training. Therefore, this paper combines the deep-learning-based stereo matching method with the traditional SIFT features. Without using additional feature enhancement, the robustness of the model for matching ill-posed regions is improved. Experimental results show that compared with the benchmark model PSMNet, the endpoint error (EPE) on the large-scale synthetic dataset Scene Flow dataset has decreased by 35.7%, and compared with ABNet [1], the breakpoint error has decreased by 18.6%. In KITTI2012 and KITTI2015, all indicators are better than those of the benchmark model, and most of the indicators are also better than those of some excellent stereo matching models. The model also has good robustness in ill-posed scenes such as strong reflection on the front windshield of a car.