2D multi-person posture estimation method combined with face detection

Abstract

The invention relates to a 2D multi-person posture estimation method combined with face detection. The method comprises the steps: firstly carrying out the research from the image content, dividing adata set into two subsets through the analysis of the number distribution of persons in the data set, then setting a corresponding branch network structure for each subset through the face detection and improved posture estimation method, reducing the false detection phenomenon of the number of people in an image by using a face detection method by using dual judgment, and when the branch networkmodel is trained, initializing model parameters by using transfer learning, and then performing optimal training by using a specific sub-data set. Compared with other methods, the method has the advantages that the coordinates of human body articulation points in the image containing the crowded or shielded scene can be more accurately predicted, and the method has higher application values for human body rehabilitation training.

Description

technical field [0001] The invention relates to the technical field of computer vision image processing, in particular to a 2D multi-person pose estimation method combined with face detection. Background technique [0002] 2D multi-person pose estimation based on static images has a wide range of applications in computer vision, such as augmented reality, virtual reality, computer animation, and 3D scene reconstruction. And the results of 2D pose estimation based on static images can be directly applied to videos. At this stage, multi-person pose estimation for 2D static images is mainly divided into two directions, namely top-down and bottom-up. Although the top-down method has high output accuracy, the running time is proportional to the number of people in the image and relies heavily on the object detection algorithm. When there is a problem with the detector, its follow-up will not be able to correctly estimate the position of the target joint point. However, the bot...

AI Technical Summary

Problems solved by technology

However, many proposed deep neural network-based methods design a unified and complex network structure for the entire training data set to estimate the human body pose. Although this improves the generalization ability of the model, it reduces the Improve the accuracy performance of the model for images containing overlapping or occluded people

There is no way to subdivide the data set for different numbers of people contained in the image and propose a corresponding solution for the crowded situation

Images