Unsupervised convolutional neural network-based monocular scene depth estimation method

Abstract

The invention discloses an unsupervised convolutional neural network-based monocular scene depth estimation method. The method comprises the following steps of obtaining a depth value of each pixel point of a target image; obtaining a camera pose value when the pixel coordinates on the target image are transferred to the next frame of image; constructing a loss function; and performing depth estimation based on an unsupervised conditional random field residual convolutional neural network scene. The problem that manual data labeling is difficult is well solved through an unsupervised method, manpower is saved, and economic benefits are improved. According to the invention, a linear chain conditional random field thought is adopted to realize feature expression of an original image. An unsupervised conditional random field residual convolutional neural network scene depth estimation model is formed by combining an unsupervised residual convolutional neural network scene depth estimationmodel. The model provided by the invention is superior to other three models in average relative error (rel) and accuracy (acc).

Description

technical field [0001] The invention relates to a method for estimating scene depth, in particular to a method for estimating monocular scene depth based on a non-supervised convolutional neural network. Background technique [0002] Computer vision is mainly a simulation of biological vision through computers and related visual sensors. People first use a camera to obtain external images, and then use a computer to convert the image into a digital signal, realizing the digital processing of the image, and finally a new discipline - computer vision, which involves many application fields including target tracking, image Classification, face recognition, scene understanding, etc. The goal of computer vision research is to enable computers to have the ability to observe the environment, understand the environment, and adapt to the environment autonomously, just like humans. [0003] However, most of the current computer vision technologies are for digital image processing. S...

AI Technical Summary

Problems solved by technology

When making a depth image data set, due to various external conditions, such as lighting, weather changes, etc., the depth sensor cannot obtain reliable and accurate image depth information, which may give the depth estimation model the accuracy of the estimated result. Influence, and this method of supervised learning has caused the problem of manual data labeling

On the other hand, as the number of layers of the convolutional neural network deepens, it may cause the problem of gradient disappearance, which brings a certain degree of difficulty to the training of the network, resulting in inaccurate results.

Images