摘要
传统的行为识别系统大多是建立在一个静态模型上,对样本特征值有非常强的依赖性,而对个别用户的行为习惯缺乏灵活适应性。基于三轴加速度计设计了一种在动态数据流环境下能够对用户行为进行增量学习的自适应性行为识别算法,该算法提出了一种新的提取加速度特征的方法,通过将三轴加速度计采集到的合成加速度数据集合看做物质,来提取物质的物理特征并训练这些物理特征用于投票分类,然后再通过增量学习来更新样本特征值,使其逐渐趋向于用户的行为习惯,从而达到更高的识别率。实验结果表明,该算法具有很高的识别率和对用户很好的适应性。
The static model used in traditional activity recognition systems greatly relies on the prior knowledge, be- ing lack of flexibility and adaptability to suit a particular user. We design an adaptive activity recognition algorithm based on triaxial accelerometer, which refines the learning model with evolving data streams. We propose a novel method for feature extraction. By taking the resultant acceleration dataset as materials, we extract their physical properties for voting classification. Furthermore, we refine the sample values by incremental learning, such that the sample values can gradually adapt to the user's activity habit and achieve a higher recognition rate. The experiment results show that our proposed algorithm owns a high recognition rate and a good adaptability to users.
出处
《传感技术学报》
CAS
CSCD
北大核心
2017年第6期909-915,共7页
Chinese Journal of Sensors and Actuators
基金
国家自然科学基金项目(61379122
61502428)
浙江省自然科学基金项目(LR16F010003
LQ15F010003)
关键词
无线传感器网络
行为识别
特征提取
增量学习
wireless sensor networks
activity recognition
feature extraction
incremental learning