基于宽接收域的实时人体姿态估计网络

Real-time human pose estimation network based on wide receiving domain

为解决人体姿态估计任务的准确率和实时性问题,提出一个卷积宽接收域、检测实时的人体姿态估计网络。构建稠密残差步进网络(dense residual steps network,DRSN),提高模型对输入图像空间信息的提取和全局特征的把握。在激活函数上,以改进的FReLU激活函数替换原始的激活函数,通过采用二维卷积的方式改变ReLU函数中的激活条件,扩大模型的接收域,关键点分类更加准确。该网络在标准MPII数据集上进行测试,在满足较高定位精度的条件下,模型在NVIDIA RTX 2080Ti GPU上的检测速度达到38 FPS,可有效解决检测实时性问题。

To solve the real-time and accuracy problem of human pose estimation task,a human body pose estimation network with wide convolutional receiving domain and real-time detection model was proposed.A dense residual step network(DRSN)was constructed to improve the model’s ability to extract the spatial information of the input image and grasp the global features.For the activation function,the original activation function was replaced using the improved FReLU activation function,and the activation conditions in the ReLU function were changed using a two-dimensional convolution method,which expanded the receiving domain of the model,and the key point classification was more accurate.This network was tested on the standard MPII data set.Under the condition of high positioning accuracy,the detection speed of the model on NVIDIA RTX 2080Ti GPU reaches 38 FPS,which can effectively solve the real-time detection problem.