Image depth estimation method and system based on CNN (Convolutional Neural Network) and depth filter

Abstract

The invention discloses an image depth estimation method and system based on a CNN (Convolutional Neural Network) and a depth filter. The image depth estimation method comprises the steps of obtainingimage depth estimation values based on the CNN, extracting local features points from the image, and establishing a minimum photometric error equation to solve the relative pose of a camera; optimizing the camera pose based on the feature points matched by polar line searching, selecting an image block by taking the feature points as a center, performing polar line searching on the image block toobtain the optimum matching, constructing a bundle adjustment equation according to the optimum matching to optimize the relative pose; and performing depth value filtering based on the camera pose,namely, filtering depth values by adopting Gaussian integration until the depth values converge. The image depth estimation method overcomes a problem of inaccurate image depth and a problem of absolute scale loss of monocular vision, and can be applied to many fields such as three-dimensional scene reconstruction, indoor positioning and augmented reality.

Description

technical field [0001] The invention belongs to the field of three-dimensional vision, and in particular relates to an image depth estimation method and system based on a CNN (convolutional neural network) and a depth filter. Background technique [0002] Most common images in reality are color images. Color images are obtained by compressing a three-dimensional scene into a two-dimensional plane. The depth information is lost during the imaging process, and the loss of depth information makes many visual tasks difficult. For example, Due to the lack of depth values, the reconstruction of 3D scenes will be difficult. Therefore, it is of great significance to recover image depth values ​​from color images. At present, the mainstream image depth acquisition methods are divided into three categories. One is to obtain the depth value through special hardware equipment, mainly RGB-D cameras, and its principle is generally structured light or time-of-flight method. Used in robot...

AI Technical Summary

Problems solved by technology

The disadvantage is that due to its computational complexity, dense depth reconstruction is generally done offline, which takes a long time

Moreover, the motion recovery structure method has strong flaws - the loss of absolute scale and the problem that pure rotation cannot be calculated

[0005] It can be seen that due to various reasons, the existing image depth estimation technology still has problems such as narrow application range, low precision, and low efficiency.

Images