Multi-person posture estimation method based on human body anchor point set and perception enhancement network

Abstract

The invention provides a multi-person posture estimation method based on a human body anchor point set and a perception enhancement network. The method comprises the steps of 1 constructing the perception enhancement network, and carrying out the network training and tuning through a training data set; and 2 for a to-be-tested picture, inputting the picture into the perception enhancement networkto carry out detection of a joint detection heat map and an anchor point embedding heat map, and then carrying out a post-processing process of an algorithm based on the two types of heat maps to obtain a multi-person posture estimation result. The method has the advantages that based on the thought of human body division, the multi-person posture estimation task is divided into the posture estimation subtasks of the upper half body and the lower half body, so that the modeling difficulty of the detection task can be effectively reduced; and a perception enhancement network is constructed by combining an attention mechanism and a feature fusion strategy, so that the feature extraction capability of the neural network model can be fully mined, and finally, a more accurate multi-person posture estimation result is obtained.

Description

technical field [0001] The present invention relates to the technical field of human pose estimation, in particular to a multi-person pose estimation method based on a human body anchor point set and a perception enhancement network. Background technique [0002] Human pose estimation technology is a basic work in the field of computer vision to study human behavior, and its goal is to detect human joints from a single RGB image. This research can be widely applied to a variety of higher-level visual tasks such as action recognition, human body tracking, and human re-identification. Since there are often many people in the actual scene, how to solve the pose estimation task of multiple people has gradually received more attention in recent years. [0003] Existing multi-person pose estimation methods can be mainly divided into two categories: top-down and bottom-up. Since the bottom-up method only needs to use a single neural network for a forward pass, it often has a lowe...

AI Technical Summary

Problems solved by technology

This invention can promote the network to learn discriminative robust features for pose estimation through an adversarial learning mechanism. However, this method also faces the same problem, that is, in some complex difficulties such as sports scenes, it may be due to joint occlusion or motion blur. and other problems, it is impossible to achieve effective search and matching between joint points

[0005] In the previous research on multi-person pose estimation algorithms, the whole human body was simply used as the target of network detection and predicted in the same branch. Due to the difficulty of coming, the algorithm is easy to cause joint point prediction errors in difficult scenes such as occlusion and motion blur, which in turn leads to wrong pose estimation results

Images