基于分组残差联合空间学习的人体活动识别

Human activity recognition based on grouped residuals combined with spatial learning

针对基于多传感器的交互性人体活动识别任务,提出了一种基于窗口预处理和分组残差联合空间学习的多传感器交互性活动识别算法。首先,针对多传感器交互性活动数据预处理过程中的滑动窗口处理方式对人体活动识别的影响进行了实验分析和实验对比,包括不同滑动窗口大小和覆盖率等;其次,基于多传感器的交互性活动识别的窗口预处理结论,利用分组残差联合空间学习进行活动识别与分类,并设计多组对比实验,分别对网络模型、损失函数和分类器等进行了优化;最后,在Opportunity活动数据集上进行对比试验,该算法性能超过了现有的大部分其他活动识别算法。实验结果验证了基于窗口预处理和分组残差联合空间学习的多传感器人体活动识别算法的有效性。

Aiming at the multi-sensor-based interactive human activity recognition task,a multi-sensor interactive activity recognition algorithm based on window preprocessing and group residual joint spatial learning is proposed.On one hand,we analyze and compare the influence of sliding window preprocessing on multi-sensor-based interactive human activity recognition,including different sliding window sizes and coverage,etc.On the other hand,based on the conclusion of sliding window preprocessing,we propose the group residual joint spatial learning for activity recognition and classification,and design multiple experiments to optimize network model,loss function,and classifier.Finally,a comparative experiment is performed on the opportunity activity recognition data set.The performance of our algorithm outperforms most of the others.It further verifies the effectiveness of the multi-sensor-based human activity recognition algorithm based on window preprocessing and grouped residual combined with spatial learning.