基于关键点提取的人体活动识别

Human Activity Recognition Based on Key Point Extraction

为改变活动识别复杂度高和活动识别率低的问题,提出了一种针对关键点提取的活动识别算法.该方法对加速度传感器采集的加速度信号进行了关键点的提取,混合隐马尔科夫回归模型(mixed hidden Markov regression model,MHMRM)对观测样本序列输出并进行建模,最大限度地通过关键点序列保留多维信号之间的特征信息,然后通过期望最大化算法(expectation maximization,EM)对模型参数进行优化并建立算法模型,使用维特比算法分割数据最终状态.使用不同活动顺序的测试集对算法性能进行测试,包含了站立、坐、躺、步行、上楼、下楼、慢跑、跳等几类活动.实验结果表明,提出的算法以关键点抽样的方式保留数据整体特征,实现快速准确地人体活动识别,其平均识别准确率为94.06%.因此,采用此方法可有效地对人体的活动信息进行分割识别,实现对人体活动的准确检测.

An activity recognition algorithm for keypoint extraction is proposed to change the problem of high activity recognition complexity and low activity recognition rate.The method extracts the key points from the acceleration signals collected by the acceleration sensors,and the mixed hidden Markov regression model outputs and models the observed sample sequences to maximize the feature information between the multidimensional signals retained by the critical point sequences,then the model parameters are optimized and modeled by the expectation maximization algorithm.The final state of the data is segmented using the Viterbi algorithm.The performance of the algorithm is tested using test sets with different activity sequences,including standing,sitting,lying,walking,going upstairs,going downstairs,jogging,jumping,and so on.The experimental results show that the proposed method preserves the overall features of the data with crucial point sampling to achieve fast and accurate human activity recognition with an average recognition accuracy of 94.06%.Therefore,the proposed recognition segmentation method can effectively segment and recognize human activity information to accurately detect human activities.