Three-dimensional attitude classification method based on two-dimensional key points and related device

Abstract

The invention is suitable for the technical field of image processing, provides a three-dimensional attitude classification method based on two-dimensional key points and a related device, and realizes that three-dimensional attitude classification of a target object can be obtained by only needing a two-dimensional image of the target object. The method mainly comprises the steps of obtaining two-dimensional key point information in a target two-dimensional image through a pre-trained deep learning network; determining key point structure features of the target object according to the two-dimensional key point information; and speculating the three-dimensional attitude classification of the target object through the key point structure features.

Description

technical field [0001] The embodiments of the present application belong to the technical field of image processing, and in particular, relate to a three-dimensional pose classification method based on two-dimensional key points and related devices. Background technique [0002] A two-dimensional image captured by a camera or a video camera will lose the information of the depth direction, so that the three-dimensional pose of the target object cannot be directly perceived when the two-dimensional image is processed by a computer. However, in the field of security monitoring, autonomous driving, smart elderly care, aviation and airports and other industries, there are a large number of computer recognition of the three-dimensional attitude of people, vehicles, aircraft, etc. in two-dimensional images, and even the recognition of physical space attitude. Requirements, such as the spatial orientation of corridors, tunnels and roads, etc. In order to enable the computer to dir...

AI Technical Summary

Problems solved by technology

[0003] Direct estimation method: Estimate 3D key points directly from the image through the end-to-end network, requiring the input to have 3D annotations. This method has fewer constraints, but requires the input and training sets to have 3D annotations. The application of the direct estimation method has a great impact on the input camera. The requirements are high, the cost is high, and such data sets are rare and captured in a controlled environment, resulting in the trained model often not being applicable to actual scenarios

[0004] Two-stage estimation method: first obtain the two-dimensional key point information of the target object, then input the two-dimensional key point information sequence into the subsequent network, and then "upgrade" to the three-dimensional key point, that is, output the three-dimensional pose; in this two-dimensional key point information sequence To include additional information, such as 3D annotations, 2D key points at different times, 2D key points at different angles of view, etc.; in applications based on human body posture, the human body parameter model can also be input according to the 2D human body key point information, and then projected The human body model optimization method for fitting two-dimensional key points; summarizing the two-stage estimation method is to first obtain the two-dimensional key point information sequence, and then obtain the three-dimensional key point estimation. The accuracy of this method is not bad, and the effect is relatively good. The data set requirements are wide and the application range is wide, but the second stage is to upgrade from two-dimensional key points to three-dimensional key points, which requires a large amount of data, so the calculation is large and slow

Images