一种轻量化实时人体姿势检测模型研究与应用

传统的OpenPose模型在人体姿势检测方面精度较好但速度较慢。为了在保证检测精度的前提下加快检测速度、缩小模型,在传统OpenPose模型基础上,使用添加二阶项融合的残差网络提取底层特征,再通过L1范数权值对训练后的模型进行权值修剪,提出了改进型OpenPose模型。实验结果表明,在检测精度大致等同原模型情况下,模型大小缩小至约8%,参数减少近83%,检测速度提升约5倍。将改进的OpenPose模型应用到仰卧起坐体育健康测试中,结果表明该模型对仰卧起坐动作检测精度达到97%。

Research and Application of a Lightweight Real-time Human Posture Detection Model

The traditional OpenPose model has good accuracy but slow speed in human posture detection. In order to accelerate the detection speed and reduce the model on condition of the detection precision, based on the traditional OpenPose model, the residual network with second-order term fusion is used to extract the low-level features, the weights of the trained model are pruned by the L1 norm weight, and an improved OpenPose model is proposed. Experiments show that when the detection accuracy is approximately equal to original model, the model size reduces to about 8%, the parameters reduces by nearly 83%, and the detection speed increases by 5 times. The improved OpenPose model is applied to the physical fitness test of sit-ups, and the results show that the detection accuracy of the model can reach 97%.