Multi-person Pose Estimation Algorithm Study Via Layer Supervision And Channel-attention Network

In recent years,pose estimation has been rapidly developed and has become a study hotspot in the computer vision field by using deep convolutional neural network.Multiperson pose estimation is a basic and very challenging task in pose estimation,meanwhile,it is a premise of advanced vision tasks such as human motion recognition,motion analysis,and human-computer interaction.In practical application scenes,intricate backgrounds,occlusions in people,and occlusions in people between objects will cause huge interference to multiperson pose estimation.In this regard,this paper aims at the two major points in multi-person pose estimation,that is missed detection of human characters and incorrect connection of keypoints.This paper makes some improvements on the existed multi-person pose estimation algorithm.This paper selects a top-down multi person pose estimation algorithm as baseline.Firstly,according to the experimental results of target detection,the YOLOv3-spp target detector is selected which is highest score at this time,and we use DIo U as the bounding box loss in YOLOv3-spp for model training,the trained model is used as the human detector to help the pose estimator in obtaining the high quality human proposal box.Secondly,a human pose estimator with channel attention mechanism is designed.The pose estimator first enhances the resolution of feature map which is down-sampling after layer by layer,through the pixel shuffle network.Then,the attention layer assigns the different weights to the feature map of each channel to achieve the purpose in adjusting the area of interests.This measure effectively solved the ambiguous prediction of keypoints in pose estimation process.The pose estimator with the attention mechanism has more accurate positioning of keypoints than before,but it is limited to visible keypoints.If some keypoints in a human instance box are blocked,it is still difficult to make accurate estimation by using a small number of visible keypoints.Finally,this paper designs a layer supervision pose estimation based on the channel attention pose estimator.The up-sampling layer in pose estimator is replaced by deconvolution layer.Then,the layers calculate the distance between the predicted value and the ground truth value to achieve the purpose of layer supervision.It can adaptively capture the local information of keypoints and human body structure from training samples,so as to predict the positions of some invisible keypoints in the test stage.The influence of GAN super-resolution on human pose estimation is also explored.In order to verify the effectiveness of the improved algorithm in this paper,three open source datasets are used to train and test the improved algorithm.From the scores obtained in these three datasets,we can draw the conclusion that the improved multi-person pose estimation algorithm is this paper has a certain competitiveness.It also can be seen from the rendered images that the improved multi-person pose estimation is this paper can effectively solve the problems of missed human detection and the misconnections of keypoints in crowd and complicated scenes.