Human body posture estimation method using adaptive data enhancement

Abstract

The invention discloses a human body posture estimation method using adaptive data enhancement. The method comprises the following steps: constructing an active transmission network ATNet and a human body paste library; sending an original image into the constructed active transmission network ATNet, and training to obtain a transformation matrix; randomly selecting a complete person from the constructed human body paste library; combining a complete person with an original image by using a transformation matrix obtained by training of an active transmission network ATNet to form a generated graph; performing loss function DLoss calculation of a high-resolution network H on an original image human body joint point heat map heatmaps and a true value ground-truth in the generated image by using the high-resolution network H, wherein the loss function D_Loss calculation is performed on the original image human body joint point heatmaps and the true value ground-true in the generated image; taking the high-resolution network H and the active transmission network ATNet as a discrimination network D and a generation network G respectively, and transmitting a loss function D_Loss value of the high-resolution network H to the generation network G; using a high-resolution network H for performing human body posture estimation on a person in an original image, and enhancing the human body posture recognition accuracy under the condition that no extra cost is enhanced, especially for some challenging cases.

Description

technical field [0001] The invention relates to the technical field, in particular to a method for estimating human body posture using adaptive data enhancement. Background technique [0002] Multi-person pose estimation refers to identifying and locating the key points of all people in a still image, and is a fundamental research technique for many vision applications, such as human motion analysis, human-computer interaction, animation, etc. [0003] Recently, with the rapid development of deep convolutional neural network (DCNN), the task of human pose estimation has made some progress. However, these methods still produce errors in some challenging situations, such as occluded keypoints and the influence of nearby people. [0004] One of the reasons for the error of DCNN is insufficient data, especially unchallenging data. For example, if the dataset lacks samples of interleaved keypoints, it will be difficult for DCNN to get the correct answer for the interleaved keyp...

AI Technical Summary

Problems solved by technology

Faced with the increasing number of people estimating human body poses in reality, some researchers have also begun to study 3D pose data enhancement, but the current research focus is on two-stage (first 2D and then 3D) 3D pose estimation. Always stay at the prediction accuracy of the 2D data itself, so as to change the network framework, rarely consider the impact of the detection frame on the enhanced data in methods such as MASK-RCNN, resulting in poor data enhancement effect, and it is difficult to truly improve the two The effect of human body posture is explained

Images