摘要：

针对传统的用户活动识别建模方法在实时性要求下精度较低的缺点,提出一种改进的实时用户活动识别建模方法。利用已标注的传感器事件流数据建立传感器触发概率矩阵,并计算出传感器距离,作为建模的先验知识,在后续建模过程中赋予每个传感器事件不同的权重。根据传感器距离的内在含义判断活动转移发生的位置,通过概率矩阵推测上次活动作为新的特征维度来建模当前活动。在Aruba、Tulum2010和HH106 3个公开数据集上的实验结果表明,与SWMI、SWMIex等方法相比,该建模方法在精度和F1 2个指标上最大提升可超过10%。

关键词: 活动识别, 特征提取, 先验知识, 滑动窗口, 传感器距离, 互信息

Abstract:

Aiming at the shortcomings of traditional user activity recognition modeling method with insufficient accuracy under real-time requirements,an improved real-time user activity recognition modeling method is proposed.The method constructs the sensor trigger probability matrix using the labeled sensor event flow data,and calculates the sensor distance.As a prior knowledge of modeling,the sensor events are given different weights in the subsequent modeling process.According to the intrinsic meaning of sensor distance,the location of activity transfer is judged,and the current activity is modeled by inferring the last activity as a new feature dimension through probability matrix.Experimental results on the three public datasets of Aruba,Tulum2010 and HH106 show that compared with SWMI,SWMIex and other methods,the proposed modeling method can increase the accuracy and F1 by more than 10%.

Key words: activity recognition, feature extraction, priori knowledge, sliding window, sensor distance, Mutual Information(MI)

中图分类号: 

• TP393