Human body posture estimation method based on dynamic lightweight high-resolution network

Abstract

The invention provides a human body posture estimation method based on a dynamic lightweight high-resolution network, and the method proposes a dynamic lightweight high-resolution network Dite-HRNet which can efficiently extract human body key point features under multiple scales and capture spatial context information among different human body key points. Through a dynamic pyramid convolution method and an adaptive context modeling method, the problems that a network module in an existing high-resolution network is too static and space context capture is insufficient are solved, and two dynamic context modules are specially designed for the high-resolution network by using the dynamic pyramid convolution method and the adaptive context modeling method. The dynamic context modules are respectively a dynamic multi-scale context module and a dynamic global context module, and finally, the dynamic context modules with different hyper-parameter configurations are applied to different branches of a lightweight high-resolution network by fully utilizing the parallel multi-branch structure characteristic of the high-resolution network. And an efficient lightweight high-resolution network is constructed.

Description

technical field [0001] The invention relates to the technical fields of deep learning and computer vision, in particular to a human body pose estimation method based on a dynamic lightweight high-resolution network. Background technique [0002] Human pose estimation, that is, the detection of the position of important human joints or parts in images or videos, is a prerequisite task for many downstream applications in the field of computer vision technology (such as behavior recognition, human-computer interaction, video surveillance, etc.). In the application of human pose estimation, especially in real-time applications under limited computing resources and equipment conditions, we must not only pursue higher detection accuracy, but also ensure high computational efficiency. At present, most researches on human pose estimation methods use convolutional neural networks to extract and detect key feature information of human bodies. Such convolutional neural networks can be ...

AI Technical Summary

Problems solved by technology

The calculation efficiency of various operations in the network module on different input data is high or low, and is affected by factors such as the resolution of the data and the number of feature channels. However, the high-resolution network contains many feature data at different scales. Therefore, the static network module cannot make good use of the multi-scale characteristics of its parallel multi-resolution network branch structure, making it difficult to achieve optimal computational efficiency.

[0004] In addition, the traditional human pose estimation network only extracts the key point features of the human body in the image through a single repeated convolution operation. Due to the limitation of the size of the convolution receptive field, the feature map extracted in this way only uses the local part of the image. Pixel information, while ignoring the contextual relationship between distant pixels

This kind of network can only learn the pixel distribution pattern of human body parts in the local area of ​​the image. Because of the lack of global information assistance, it cannot grasp the spatial context relationship between all human body parts well, resulting in the deviation of image feature extraction.

Increasing the size of the convolution kernel used by the convolution layer can expand the receptive field range of each operation on the image, thereby capturing more image space context information, but blindly increasing the size of the convolution kernel will cause the network The computational complexity is gradually increasing, which is not conducive to the lightweight design of the network

Images