Understanding and Modeling of WiFi Signal Based Human Activity Recognition论文阅读笔记

Understanding and Modeling of WiFi Signal Based Human Activity Recognition

Wei Wang Alex X. Liuyz Muhammad Shahzadz Kang Ling Sanglu Lu
State Key Laboratory for Novel Software Technology, Nanjing University, China
Dept. of Computer Science and Engineering, Michigan State University, USA
ww

Challenge

1. 随着人的运动，信号传播的多径会产生变化，而且人身体的不同部分在做给定动作的时候会有不同的速度，应对这一挑战的方法是离散小波变化（DWT）
2. 在不同的时间不同人甚至同一个人做同一个动作会产生不同信号pattern，为了解决这个问题，作者提出了Hidden Markov Model (HMM) ，这一模型只和身体的运动速度有关，而给定的身体运动速度往往对应了给定的activity
3. CSI容易被噪声淹没，其来源包括：环境中的电磁噪声，the internal state changes in WiFi devices, e.g.,transmission rate adaptation and transmission power adaptation often introduce impulse and burst noises in CSI values .而且，因为滤波器所需要的频率过高，同时信噪比比较低，所以传统滤波方法不可行。为此，作者采用了principal component analysis (PCA)
4. 避免carrier frequency offset (CFO) 的干扰，作者使用CSI signal power
5. 如何自动给人的动作切段，方法是使用PCA中提取的特征向量——We capture the smoothness of the eigenvector by calculating its high-frequency energy and compare it to a dynamically adapting threshold to detect start and end.

Contribution

1. propose the CSI-speed model and the CSI-activity model to quantify the correlation between CSI value dynamics and a specific human activity.
2. propose a set of signal processing techniques,
   such as PCA based denoising and DWT based feature extraction, for human activity recognition based on the CSI-speed model and the CSI-activity model.

UNDERSTANDING WIFI MULTI-PATH

1. H(f; t) is the complex valued channel frequency response (CFR) and the time-series of CFR values for a given antenna pair and OFDM subcarrier is called as CSI stream .
   Thus, there are 30 × NT x × NRx CSI streams in a time-series of CSI values.
2. 经典公式的起源
   
3. 对于本文而言，因为是使用CSI siginal的能量，所以作者建立了能量和运动速度的联系，发现能量的大小是运动速度的函数。推导过程如下：

MODELING OF HUMAN ACTIVITIES

首先通过时频分析确认了走路、摔倒、坐下三种不同的速度确实可以和不同的频率相对应。
然后通过隐马尔可夫模型构建的有限状态机通过不同动作反应的频谱特征串联起人体动作的行为识别。
在讨论中，论证了对高低速运动的普适性，对身体不同部位的识别，而对于多人运动则需要多套收发设备。

PCA BASED CSI DENOISING SCHEME

First, CARM collects CSI values and removes the noises in the measurements.

噪声来源分析
transmission power changes, transmission rate adaptation, and internal CSI reference level changes. These internal state transitions result in high amplitude impulse and burst noises in CSI streams.
An interesting feature of these impulse and burst noises is that their effect is highly correlated across all CSI streams, i.e., they affect samples in all streams at the same time.
使用PCA，分四步
With PCA, we can track the time-varying correlations between CSI streams, and optimally combine them to extract principal components of CSI streams.

1. Preprocessing: In this step, CARM first removes the static path components from each CSI stream by subtracting the corresponding constant offsets from the streams.
   减掉静态部分

2. Correlation estimation: CARM calculates the correlation matrix
   计算相关矩阵，即HxH‘

3. Eigendecomposition: CARM performs Eigendecomposition of the correlation matrix to calculate the eigenvectors.
   计算特征向量

4. Movement Signal Reconstruction: In this step, CARM constructs the principal components using the equation hi = H × qi, where qi and hi are the ith eigenvector and the ith principal components, respectively. CARM discards the first principal component h1 and retains the next five principal components to be used for feature extraction. 构建，同时扔掉第一个

Second, CARM extracts human movement features from the denoised CSI values using DWT.

人行为的两大特征：duration and frequency

使用DWT提取frequency信息
同时，作者还申明了：The advantage of DWT compared to STFT is as follows: First, DWT has nice tradeoffs in time and frequency resolutions. DWT naturally groups frequencies that differ by several orders of magnitude into a few levels so that both high speed movements and low speed movements can be captured. Second, DWT reduces the size of data so that the classification algorithm can run in real time.
具体的提取方式节选如下：
To extract features for classification from a sample of an activity, CARM applies DWT to decompose the PCA components into 12 levels that span the frequency range from 0.15Hz to 300Hz. The
DWT results of the five PCA components are averaged to capture the movement information present in different PCA components. From the output of DWT on each 200ms interval, CARM extracts a 27 dimensional feature vector that includes three types of features. 1). The energy in each level, which represents the intensity of movement in each speed range. 2). Difference in the energy of each level between consecutive 200ms intervals, which represents the rate of change of speed of a multi-path for the activity. 3). Estimated torso and leg speeds using the percentile method introduced in Doppler radar
最后宣称了环境无关。

Third, CARM trains an HMM model for each activity and uses the CSI-activity models to recognize activities in real time.

1. 使用了Baum-Welch algorithm

2. 行为segment的处理：Activity Detection: To detect the start and end of an activity, CARM monitors the second eigenvector q2 and the corresponding principal component h2. Our activity detection method is based on two key observations. First, in the absence of an activity, the eigenvector q2 varies randomly over neighboring subcarriers because CSI streams contain uncorrelated values; whereas, in the presence of an activity, the CSI streams become correlated and q2 varies
   smoothly over neighboring subcarriers. Second, in the absence of an activity, the principal component h2 has smaller variance; whereas in the presence of an activity, it has higher variance.

3. 因为要根据噪声水平动态调整判断行为开始结束的判定阈值，所以使用了 Exponential Moving Average (EMA)算法

4. 最后的识别部分Activity Recognition: CARM identifies the activity in following four steps. First, it takes all CSI values between the start and end times and denoises them using the PCA based denoising method described in Section 5.4. Second, from every 200ms interval between the start and end times, it extracts a 27 dimensional feature vector as described in Section 6.1. Third, it uses dynamic programming to calculate the likelihood of each HMM generating this sequence of feature vectors . Finally, the model with the highest likelihood identifies the activity.