A Binocular Depth Estimation Method for Driving Scenes Overcoming Occlusion Effect

Abstract

The invention discloses a binocular depth estimation method for driving scenes that overcomes the occlusion effect. This method constructs an end-to-end self-supervised deep neural network model, input images of left and right views of the driving scene, and output disparity maps corresponding to the left and right views. This method uses the geometric constraint relationship between the input and output of the deep neural network model to train the model. It does not need to obtain data samples with labeled information, but only needs to use the binocular camera system to obtain the image pair of the left and right views, which greatly simplifies the workflow. The economic cost is saved, and the model can be trained on images of more types of scenes. The binocular estimation method designed by the present invention can effectively overcome the problems of repeated pattern textures and occlusion of front and rear objects in the scene, and can obtain a higher-precision depth image.

Description

technical field [0001] The invention relates to the fields of machine vision and automatic driving, in particular to a binocular depth estimation method for driving scenes that overcomes occlusion effects and utilizes self-supervised deep learning technology. Background technique [0002] With the advancement of artificial intelligence technology, autonomous driving has been widely studied in academia and industry. As an important part of autonomous driving technology, binocular depth estimation has always been a research hotspot. The binocular depth estimation is based on the binocular camera, shoots two left and right views, obtains the corresponding disparity map from the left and right views, and then calculates the depth image according to the binocular camera parameters. [0003] The traditional binocular depth estimation uses a stereo matching method to find matching corresponding points in the left and right views. However, due to the existence of the occlusion effe...

AI Technical Summary

Problems solved by technology

Although this self-supervised training mechanism can avoid the trouble of collecting real scene depth information, due to the existence of occlusion effects, it is impossible for the reconstructed view to be completely consistent with the original view, which will lead to a decrease in global accuracy.

A network model trained in a self-supervised manner will be much less accurate than a network model trained in a supervised manner

Images