面向边缘设备的轻量级OpenPose姿态检测模型研究

A lightweight OpenPose posture detection model for edge devices

为了在低算力的边缘设备上使人体行为识别网络兼顾实时性和识别效果,提出了一种改进的轻量级OpenPose姿态检测模型。使用移动网络替换原主干特征提取网络,在特征提取网络的浅层使用倒置残差结构,减少浅层网络的运算量,在网络深层引入卷积块注意力模块,调整深层特征信息的权重,并将浅层网络特征与深层网络特征融合后送入卷积网络进行骨骼关键点的拼接,有效融合浅层和深层的特征信息。在COCO数据集上的验证结果表明:改进模型与原模型相比,正确关键点百分比提升了2.8%,平均精度提升了2.0%。使用改进后的模型作为预训练模型在行为数据集上标记骨骼关键点用作分类训练,将完成分类训练的模型部署在边缘设备上,在边缘设备运行速度略微降低的情况下,通过改进后的模型进行人体行为识别的准确率达到96.4%,有效实现在边缘设备上的姿态检测和人体行为识别。

In order to achieve real-time human behavior recognition on edge devices with low computational power while keeping a balance between real-time performance and recognition effectiveness,we proposed an improved lightweight OpenPose posture detection model.The model replaces the original backbone feature extraction network with mobile networks,uses inverted residual structures in the shallow layers of the feature extraction network to reduce computational complexity,introduces convolutional block attention module in the deep layers to adjust the weights of deep feature information,and fuses shallow and deep feature information after merging them before feeding them into the convolutional network for the concatenation of skeleton keypoints.This effectively integrates both shallow and deep feature information.Validation results on the COCO dataset show that compared to the original model,the improved model achieves a 2.8% increase in correct keypoint percentage and a 2.0% increase in average precision.Using the improved model as a pre-trained model,skeletal keypoints are labeled on a behavioral dataset for classification training.When deploying the trained model on edge devices,even with a slight decrease in operating speed on edge devices,the accuracy of human behavior recognition reached 96.4%,effectively realizing posture detection and human behavior recognition on edge devices.