一种针对视频的人体姿态估计的加速算法

An Accelerated Algorithm of Human Pose Estimation for Video

为了解决视频的高性能人体姿态估计算法参数量和计算量庞大导致的推理速度慢的问题,提出了基于高分辨率网络(HRNet)的人体姿态估计改进算法。该算法在检测过程中采用隔帧检测和去抖动的优化处理,优化人体检测流程;针对姿态估计网络,使用ShuffleUnitV2组件对HRNet重新设计得到了S-HRNet,提高网络的利用率。实验结果表明:在公开数据集COCO训练集上,改进算法的总推理时间为356 ms,而原始算法总推理时间为992 ms,有效地提高推理速度。改进后的算法解决了原有的HRNet模型参数量大、推理速度慢的问题,同时也保持了一定的性能,为实际部署提供了一个适合的算法。

In order to solve the problem of slow inference speed caused by the large number of parameters and computation of high performance human pose estimation algorithm for video,an improved algorithm of human pose estimation based on High Resolution Network(HRNet)is proposed.The algorithm uses interframe detection and de-jittering optimization in the detection process to optimize the human detection process.For the pose estimation network,the S-HRNet is obtained by redesigning the HRNet using ShuffleUnitV2 component to improve the utilization of the network.The experimental results show that the proposed accelerated algorithm for human pose estimation has a total inference time of 356ms in the public data set COCO training set,while the original algorithm has a total inference time of 992ms,which effectively improves the inference speed.The improved algorithm solves the problems of large parameters and slow inference speed of the original HRNet model,while maintaining a certain performance,providing a suitable algorithm for actual deployment.