Monocular depth estimation method, apparatus, terminal, and storage medium

Abstract

The embodiment of the invention discloses a monocular depth estimation method, a device, a terminal and a storage medium. The method comprises the following steps: obtaining monocular video; determining a semi-dense depth map corresponding to key image frames in the monocular video according to a preset reconstruction algorithm; using the key image frame and the semi-dense depth map as inputs of apreset neural network model, and determining the dense depth map corresponding to the key image frame according to the output of the preset neural network model. The technical proposal of the embodiment of the invention effectively combines a preset reconstruction algorithm and a preset neural network model, thereby obtaining a dense and high-precision depth map.

Description

technical field [0001] Embodiments of the present invention relate to image processing technologies, and in particular to a monocular depth estimation method, device, terminal and storage medium. Background technique [0002] Depth estimation has important research significance in the fields of autonomous driving, robot obstacle avoidance, and augmented reality. The depth estimation method can be used to estimate the depth information of each pixel in an image to obtain a depth map of the image. In the prior art, depth information can be obtained directly through multiple sensors such as lidar and depth cameras, but these sensors have a certain volume and high cost, which limits the application range of depth estimation. It can be seen that if only one camera is used for depth estimation, the application scenario can be greatly simplified. [0003] Currently, there are two methods for depth estimation using a monocular camera. The first is a method of monocular depth esti...

AI Technical Summary

Problems solved by technology

[0004] However, for the first estimation method, although a dense depth map can be predicted, because the neural network is limited to semantic understanding and there is no geometric measurement information, the accuracy of the depth map predicted by the neural network is poor. Lack of trustworthiness and low generalization ability of the network

For the second estimation method, a more accurate depth value can be obtained by solving multi-view geometry, but whether it is a SLAM system based on a feature point method such as ORB (Oriented FAST and Rotated BRIEF)-SLAM algorithm, or a SLAM based on a direct method Systems such as the large-scale monocular LSD (Large-scale Direct)-SLAM algorithm can only obtain depth information of a small number of feature points or high-gradient points, so that dense depth maps cannot be obtained, and only sparse or semi-dense depth maps can be obtained.

It can be seen that the existing monocular depth estimation method cannot obtain a dense and high-precision depth map

Images