Human pose estimation based on directional image fusion

Abstract

The invention discloses a human pose estimation based on directional image fusion, which comprises the following steps: extracting features to obtain high-level features; performing preliminary learning detection of key point location map and detection of directional image to obtaina position heat map and a directional heat map. The predicted position heat map, direction map and high-level features are fused to get more accurate position heat map. By fusing the direction information, the invention improves the prediction precision of the key points, and compared with the previous network structure, the proposed depth convolution network can obtain more human posture information. The invention can effectively solve the attitude estimation under the occlusion problem. For the occluded key points, when the position heat map is basically correct, even if the position network of the first stage cannot find the key point position well, the back fusion network can accurately predict the keypoint position, and the robustness is better. The network structure parameters of this method are much smaller, and the requirement of display memory is not high in the training phase, so the trainingis easier to converge.

Description

technical field [0001] The invention belongs to the technical fields of computer vision and image processing, and relates to a human body posture estimation technology, in particular to a human body posture estimation method for merging direction diagrams. Background technique [0002] With the popularity of smart phones, tablet computers and other devices, images and video data for human body activities are generated all the time. How to make computers automatically understand human body movements is also very important. Its application fields will be very wide, such as human computer interaction, intelligent monitoring, etc. Computers can efficiently and automatically understand human movements, which will have a profound impact on the entire society. In this context, human pose estimation was proposed. The purpose of human pose estimation is to detect the pose behavior of the human body, and to obtain digital and presentable pose information through computer learning an...

AI Technical Summary

Problems solved by technology

This method has certain robustness, but it is difficult to detect the key points of the part that are occluded or the key points of the part that tend to be consistent with the background color.

[0004] To perform human pose estimation, many technical problems need to be solved: (1) The occlusion problem is the most difficult problem in pose estimation

In the existing methods, it is basically to directly detect the position heat map, let the deep convolutional neural network implicitly learn the relationship between the key parts of the human body, which makes it difficult to detect the occluded human body parts, and also slows down the training speed

(2) Fusion of the foreground image and the background image, that is, the color of some key point areas of the human body is basically the same as that of the adjacent area in the background image, which leads to the inability to accurately estimate the key point position during key point detection, resulting in a large detection error

(3) It is difficult to distinguish between key points

For example, the left and right feet are indistinguishable in a sideways position

Images