基于轻量化二维人体姿态估计的小样本动作识别算法

A Small-sample Action Recognition Algorithm Based on Lightweight Two-dimensional Human Posture Estimation

关键词:  时序数据挖掘  动作识别  人体目标检测  人体姿态估计  动态时间规整

Key words:  temporal data mining  action recognition  human target detection  human posture estimation  dynamic time warping

摘要:

动作识别是近年来时序数据挖掘领域的研究热点，具有广泛的应用前景。但是现阶段基于深度学习的动作识别算法需要大量的标记训练数据集，存在泛化性差、实时性差、场景受限的问题。为解决这些问题，本研究设计一种基于轻量化二维人体姿态估计的小样本动作识别算法。该算法基于YOLOv5算法构建轻量化的人体检测器HYOLOv5。基于轻量化二维姿态估计模型Lite-HRNet设计人体姿态特征描述算子，有效地去除背景对人体动作特征的干扰。为有效度量时序人体姿态特征描述算子间的相似度，本研究提出基于动态时间规整的人体姿态特征距离度量，并在此基础上设计基于类别中心选择的动作模板匹配算法。该算法通过少量的动作视频构建动作特征模板库，利用动作模板匹配算法可实现多类动作视频的精准识别。为验证算法，本研究在COCO 2017的Humans数据集上对HYOLOv5进行测试，人体检测识别精度mAP@0.5：0.95可达50.7%。基于10种动作视频数据进行测试，结果表明，本研究所提算法可有效地识别视频序列中的姿态，在每个动作仅包含4个训练数据的情况下，动作识别准确率均可达到91.8%。

Abstract:

Action recognition is a research hotspot in temporal data mining in recent years with a wide range of application prospects.However, the action recognition algorithm based on deep learning at the present stage requires a large number of labeled training datasets, which has the problems of poor generalization, poor real-time performance and limited scene.In order to solve these problems, this study designs a small-sample action recognition algorithm based on lightweight two-dimensional human posture estimation.The algorithm builds a lightweight human detector HYOLOv5 based on the YOLOv5 algorithm.Based on the lightweight two-dimensional posture estimation model Lite-HRNet, a human posture feature descriptor is designed to effectively remove the interference of background on human action features.In order to effectively measure the similarity between temporal human posture feature descriptors, this study proposes a human posture feature distance measurement based on dynamic time warping, and designs an action template matching algorithm based on category center selection.The algorithm constructs a template library of action features through a small number of action videos, and uses the action template matching algorithm to achieve accurate recognition of multiple types of action videos.To verify the algorithm, this study tested HYOLOv5 on the COCO 2017 Humans dataset, and the human detection recognition accuracy of mAP@0.5:0.95 could reach 50.7%.Based on 10 kinds of action video data, the results show that the proposed algorithm can effectively identify the posture in the video sequence.When each action contains only 4 training data, the accuracy of action recognition can reach 91.8%.