ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2017, Vol. 54 ›› Issue (12): 2731-2740.doi: 10.7544/issn1000-1239.2017.20150462

• 网络技术 • 上一篇    下一篇



  1. 1(中国科学院计算技术研究所 北京 100190); 2(中国科学院大学 北京 100049) (
  • 出版日期: 2017-12-01
  • 基金资助: 

A Time Window Based Lightweight Real-Time Activity Recognition Method

Dong Lihua1,2, Liu Qiang1, Chen Haiming1, Cui Li1   

  1. 1(Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190); 2(University of Chinese Academy of Sciences, Beijing 100049)
  • Online: 2017-12-01

摘要: 利用手机或可穿戴设备实时识别人的运动状态,有助于人们及时了解自身状况,进行科学的锻炼.现有高准确度运动识别算法大都具有较高的计算代价和存储代价,难以直接移植到手机和可穿戴设备上,且这些算法难以根据用户习惯校正识别模型.提出了一种基于时间窗口的轻量级实时运动识别算法EasiSports,利用手机或可穿戴设备中的加速度传感器,在多种情况下利用k-means聚类等方法在设备本地建立用户个人运动识别模型,使用SVM分类器实时识别坐、步行、跑步、上楼梯、下楼梯5种状态,计算量较小,适用于手机和可穿戴设备平台.实验表明:该算法对前述5种状态的识别准确度可达到87.45%,识别算法运行时间相较其他算法可获得30%以上的性能提升.

关键词: 运动识别, 智能手机, 可穿戴设备, SVM分类器, k-means聚类

Abstract: Lack of exercise or excessive exercise would damage our body, so real-time human activity recognition using smart-phones or wearable devices can keep people aware of their physical status and contribute to proper exercise. However, most of existing high-precision activity recognition algorithms cannot directly apply to smart-phones or wearable devices due to their high computational cost and large storage request. Meanwhile, these algorithms cannot update recognition model according to user behavior. We propose a time window based lightweight real-time activity recognition method (EasiSports) which uses SVM to recognize five kinds of human activity status including sitting, walking, running, walking upstairs and downstairs with one accelerate sensor and fair low computational cost. Our method can set up personalized activity recognition model in many cases using k-means clustering and preset data to further improve the accuracy and reduce computational cost. Experiments show that our method can achieve an accuracy of 87.45% in recognizing five kinds of human activity status above mentioned. The performance of our algorithm can improve more than 30% compared with other algorithms.

Key words: activity recognition, smart phone, wearable device, SVM classifier, k-means clustering