WiFi-Enabled Smart Human Dynamics Monitoring

Related work Device-based methods Device-free methods (without CSI) limited by requiring the users to carry the appropriate equipments Device-free methods (without CSI) limited by high deployment costs and privacy concerns Device-free methods (with CSI) focus on one specific aspect or coarse-grained 基于一些设备和传感器的方法 需要专用的设备 使用RSSI或者图片等信息(拍照，摄像)来获取 部署话费高或者拍照涉及到隐私保护 使用CSI来获取 要么方面比较单一 要么精度不高

This work A fine-grained comprehensive view of human dynamics using existing WiFi infrastructures Participant number estimation Human density estimation Walking speed and direction derivation 使用wifi做一个细致全面获取动态数据的系统 人数量的估计、密度的估计、行走速度和方向的估计(一个人)

CSI Fundamental In a multipath wireless environment, the received wireless signal can be expressed as: y and x = are the receiving and transmitting signals in vector form H = and n are the channel response and the ambient noise Estimation of H represents the channel state information (CSI) readily available from many commercial wireless devices three major factors: amplitude attenuation, phase shift and the propagation delay 公式具体解释： Ai 是第i个路径的复值幅度衰减 ai * e jQi e –jwt 是第i个路径的相位偏移 t 传播时延 = l/c l:第i条路径 c 是光速

SYSTEM DESIGN Data Preprocessing module Three functional modules filter out high frequency noise remove the subcarriers that are not sensitive to capture human dynamic characteristics Three functional modules (our system takes time-series of raw CSI readings as the input and each CSI measurement contains 30 Nt × Nr , matrices, where Nt and Nr are the number of antennas on transmitter and receiver, respectively.)

Participant Counting Investigate and determine the effective features monotonic relationship Describe the monotonic relationship via a non-linear model multiple dimensions(different links, frequency bands, channels) can be generalized and applied to other indoor environments (只需要采集晚上或者早上空房间的CSI，更新参数即可)

Participant Counting Feature Extraction CSI amplitude variance CSI amplitude range CSI amplitude mode Entropy of HIP (i.e., phase difference). (从Mimo支持的wifi设备的两跟天线间CSI差异提取有效特征) 从CSI数据的相位和幅值信息提取出四个特征 幅度方差、幅度范围、幅度模式、HIP熵

Participant Counting Feature-based Non-linear Regression logistic bounded exponential function: use inverse function (单一feature) The final estimate on the number of participants is obtained by minimizing the total estimation errors produced by these N features as follow: 逻辑有界指数函数 五个场景求解参数 得到单一特征的模型 Yfi第i个特征输出 wi实验测定 拓展到其他环境。a,b决定模型的形状、偏移 c,d描述形状曲率，在不同环境是稳定的 测定a，b即可

Participant Counting Multiple Links is the output from the i-th link Multiple Channels is the output from the i-th channel 求平均 两条链路 Wl = 0.5 四个信道 Wc = 0.25

Human Density Estimation Distribution Analysis over Multiple Subcarriers Earth Mover Distance(EMD) Calculation Profiles built from a fixed number of people would be enough 评测在哪一个块区域比较密集，比如四块，哪块人多 首先根据密度分布构建CSI在多重载波上的幅度方差分布图 得到CSI计算幅度方差分布，EMD匹配 EMD用于评估两个概率分布的相似度的一个方法 sensitive 载波： 邻居拥有较低方差 位置独立的，与人数无关，只是值有所不同

Speed and Direction Derivation Phase Difference Information Extraction Total Harmonic Distortion (THD)-based algorithm(speed) FFT-based algorithm(direction)

THD-based algorithm(speed) Predefined threshold to determine static or walking period given the spectrum derived from the time series of CSI relative phases, the power level in lower frequency band decreases as the working speed increases THD analysis to perform walking speed estimation THD分析导出输入信号的周期图，找到基带信号，得到一些列新的csi的值

THD-based algorithm(speed) e = h * w monotonic relationship with walking speed each peak of the fundamental signal stands for the scenario that the person walks across the LOS between the transmitter and receiver

FFT-based algorithm(direction) Segment CSI measurements using a fixed time window (3 ∼ 5 seconds) Define the frequency distribution as: only examine the dominant frequency bin e(1) 三个方向： 0,45,90(某一链路) e(n)表示第n个输出fn进过fft变化的频率分布 (采样率50hz 5点fft变换 5段 10hz的带宽) e(1)受影响最大 D图 能量分布 和link2角度

FFT-based algorithm(direction) CSI frequency distribution and variance on two perpendicular wireless links as a combined feature vector Trains a support vector machine (SVM)model with Gaussian kernel Classifies the walking direction based on the pre-built walking direction profiles above(three walking directions (i.e., 0◦, 45◦, 90◦)) 某一方向的高速行走导致低方差值 低速导致高方差的分布

EVALUATION The Number of Participants Estimation Non-linear Model Construction 较好的单调性

EVALUATION The Number of Participants Estimation semi-supervised Learning Approach 某个建好的模型拓展到其他室内 在A位置建好的模型在A，B，C表现 80%准确率

EVALUATION The Number of Participants Estimation data Fusion(multi-links & multi-channels) 5G表现好 2.4G干扰大一些 (c)(d)without channel combination prove that using more channels improves the performance of our system up to 8%. (We observe that without channel combination, the system performance at three different locations degrades to 70%, 70% and 86% at 2.4GHz and 74%, 81% and 87% at 5GHz.)

EVALUATION Human Density Estimation 从A建好的profiles应用到ABC三处的准确率 (indicate that profiles built from the training rooms can be extended to other rooms since the pattern of the CSI distribution is preserved) 不同人数建立的profile的EMD计算相似度 成功预测 R1对应R1 Successfully predicts the people density region using profiles constructed from different numbers of people

EVALUATION Walking Speed and Direction Derivation walking speed

EVALUATION Walking Speed and Direction Derivation walking direction 分类结果 accuracy of 90% B 训练数据 (We show the walking direction classification results as a confusion matrix in Figure 15(a). .e results show that our walking direction estimation method can achieve an average accuracy of 96.6% for the three walking directions)

谢谢聆听~ We will explore the walking speed and direction estimation under the multi-people scenario in our future work.