E-eyes: Device-free Location-oriented Activity Identification Using Fine-grained WiFi Signatures 笔记

E-eyes: Device-free Location-oriented Activity Identification Using Fine-grained WiFi Signatures

Yan Wang, Jian Liu, Yingying Chen, Marco Gruteser, Jie Yang, Hongbo Liu
MobiCom’14, September 7-11, 2014, Maui, Hawaii, USA

Contribution

1. We show that the channel state information (CSI) from offthe-shelf 802.11n devices can be utilized to identify and distinguish in-place activities inside a home with a much smaller set of transmitting devices than in previous device-free localization solutions.
2. We develop a monitoring framework that can run on a single WiFi AP with its connected devices and use the associated profile matching algorithms to compare amplitude profiles against those from known activities.
3. We explore dynamic profile construction. The profile can be adaptively updated to accommodate the movement or replacement of wireless devices (e.g., laptop or smartphone) and the day-to-day profile calibration.
4. We conduct extensive experiments in two different-sized apartments over a 4-month time period to demonstrate that a single AP with 3 connected devices can accurately distinguish 8 walking activities between rooms (20 rounds each), 9 daily activities (50 rounds each), and over more than 100 rounds
   of other activities with an average detection rate of over 96% and an average false positive rate less than 1%. With only one device, the detection rate can still achieve around 92% with a similar false positive rate.
5. We show through experiments how the trend to wider channels (e.g., 802.11ac) will further enhance recognition since itallows measurements over many additional subcarriers.

Limitations

first, our system E-eyes is mainly designed for and tested with a single occupant at home, which is an important scenario, such as in an aging-in-place environment. It may be extended to multiple persons, however this would require a much larger set of profiles covering different combinations of activities. Second, E-eyes was tested without pets or other movements at home. While the effect of small pets is probably negligible, larger pets may pose additional signal processing challenges. Third, E-eyes relies on a relatively stable environment (e.g., no furniture movement). While it could detect such environment changes, each would trigger an activity profile updating procedure, which may require user input.

DESIGN OF E-EYES

Intuition

现阶段两个趋势：

1. 随着越来越多的设备可以接入AP，提供了更多的链路；
2. 提供了更细粒度的信道信息和带宽；

amplitude measurements on each subcarrier will provide many uncorrelated combinations of the received multipath components, which increases the likelihood that some are affected by a small movement.

The insight is that since an activity involves a series of body movements during a certain period of time, the distribution of CSI amplitudes is a desirable channel statistic that can capture unique characteristics of activities in both time and frequency domains.

本文观点就是，既然使用CSI可以拿到更加细粒度的信息，那么原来使用大量设备、统计RSS形成的感知“mesh”的方式就可以用少量设备通过多径叠加来取代。

Challenges

Profile Uniqueness and Robustness 因为CSI受干扰太大，所以很那将信号特征匹配到某一个动作上
Algorithm Generality 不同动作会在不同方面展现特征，例如走路是持续一段时间不同动作，并涉及位移，而洗碗却是原地重复的动作
Profile Generation 因为环境可能会变，所以特征可能会不一样

System Overview

1. 输入：静止设备上一段时间序列上子载波的振幅信息
2. 使用低通滤波去除异常值和人为影响
3. 行为识别包含两部分，先识别是行走和原地活动——a walking activity causes significant pattern changes of the CSI amplitude over time, since it involves significant body movements and location changes. An in-place activity (such as watching TV on a sofa) only involves relative smaller body movements and will not cause significant amplitude changes but present certain repetitive patterns. 使用的是动态自适应的方差阈值方法进行区别这两种行为，同时，变化的方差也用来进行行为的分段
4. 对于行走，使用Multiple-Dimensional Dynamic Time Warping (MD-DTW) technique, which can align a trace with larger CSI changes to the profile while correcting for differences in speed。
   而对于原地活动，使用Earth Mover Distance (EMD) 对CSI distributions (i.e., histograms) 进行识别
5. 每到一个环境需要重新找pattern，需要用户自己标注自己的行为cluster，环境变动以后也需要更新

ACTIVITY IDENTIFICATION

Coarse Activity Determination

1. 根据CSI振幅C和P的方差计算累计方差（每个时刻算一个，结合所有子载波的数据）
2. 通过一个阈值τ 衡量V ，如果超过阈值就是行走，否则就是原地运动
3. 这个方法也被用来判断动作的起始和结束

In-place Activity Identification

简言之就是使用EMD判断是不是达到判定为某一动作的阈值，但是还有一些细节需要注意；
An alternate way to determine whether the testing CSI measurements correspond to a known activity or not is to use an outlier detection method, such as the median absolute deviation (MAD) [27], to examine whether the resultant minimum EMD distance is within a range. To determine the range, an EMD distance pool containing the minimal EMD distances of previous successfully identified activities is needed in the profiles. We note that our system can also recognize the same in-place activities occurring in different locations by comparing the testing CSI measurements to a set of CSI profiles constructed when the same activities occur in different locations. In this case, the profile for an activity is a set of CSI profiles instead of a single CSI profile, and the testing CSI measurements are determined to contain the activity if it has the minimum EMD distance to any of the CSI profiles belonging to the activity profile.

Walking Activity Tracking

1. 行走路线的识别用了一个技巧，即只识别人经过了哪一个doorway，就知道人从哪一个地方到了其他地方（By identifying the doorway the person moves through, our system can determine a walking activity in high level without requiring extensive profiling of paths that have less meaningful starting and ending locations. ）
2. 行走路径区分：使用Dynamic Time Warping (DTW) 进行匹配，当然了，为了使用多载波信息，To perform multidimensional sequence alignment, our system employs Multi-Dimensional Dynamic Time Warping (MD-DTW) [33], in which the vector norm is utilized to calculate the distance matrix according to:
   
3. Doorway区分：使用滑窗，在每一段内使用EMD进行判断

Extension to Wider-band

更大带宽的好处就是更高的数据采样和更多子载波

IMPLEMENTATION

1. the distribution of CSI amplitude to profile in-place activities and the sequence of CSI amplitude for walking activities.

2. non-profiling clustering, in which E-eyes first utilizes a semi-supervised approach to cluster the daily activities from the collected CSI measurements, and then label each activity to produce CSI profiles. After clustering, the CSI measurements from the same type of activities are clustered together. Once significant changes of profiles are detected, E-eyes utilizes users’ feedback (i.e., user labels each new cluster returned by non-profile clustering) to perform adaptive profile updating. The non-profiling lustering is also utilized to construct CSI profiles when our system starts without any CSI profile.
   简单来说，就是对K个实例使用EMD作为不同实例的距离做K-means聚类

数据处理

1. 低通滤波：dynamic exponential smoothing filter (DESF) since it is an exponential smoother that changes its smoothing factor dynamically according to previous samples. The DESF can remove high frequency noise and preserve the features affected by human activities in the CSI measurements.
2. the modulation and coding scheme (MCS) index, which occasionally changes due to the unstable wireless channel in our experiments, could also influence the amplitude of CSI. In particular, we find that CSI measurements with the MCS index greater than 263 1 can make CSI measurements relatively stable in empty rooms even though it changes in such a range. Therefore, we filter out the CSI measurements with MCS value less than 263 and keep the rest of them for activity identification.

Data Fusion Crossing Multiple Links

建议看原文，公式比较好懂。