CN106066995B - A Wireless Unbound Human Behavior Detection Algorithm - Google Patents

Abstract

The invention discloses a wireless unbound human behavior detection algorithm, the purpose of which is to identify human behavior by analyzing different change patterns of channel state information, which can satisfy convenience and safety while achieving high recognition accuracy. The technical solution adopted is: use the wireless transmitter to establish a wifi field, when the user walks in the wifi coverage area or makes a certain action, it will have a specific impact on the wifi channel, and use the wireless receiver to receive the wifi signal and calculate the action CSI value, extract channel change features, use the different change features of the wifi channel when people move in the wifi field or perform certain actions, analyze the different change patterns of channel state information, and use feature extraction and classification matching algorithms to perform behavior Recognition, combining human behavior and different change modes of the channel, so as to realize the recognition of human behavior with wifi channel characteristics.

Description

A Wireless Unbound Human Behavior Detection Algorithm

technical field

The invention belongs to the fields of feature extraction, pattern recognition and behavior detection, and in particular relates to a wireless unbound human behavior detection algorithm.

Background technique

At present, with the development of science and technology and the improvement of people's living standards, the concept of smart home and virtual reality technology have developed rapidly. For example, people can control smart devices indoors through specific gestures, and achieve better game experience and human-computer interaction through physical behavior simulation operations. At the same time, people have also put forward new requirements for life monitoring technology. A health monitoring system that reflects people's physical condition by detecting abnormal behaviors such as people's sitting and sleeping positions, whether they smoke or not; a life warning system that can promptly prompt and notify medical staff and family members before the elderly or young children fall or fall. The prerequisite for the realization of the above technologies is that the system is required to accurately detect and recognize people's physical behavior.

At present, the methods for realizing human behavior detection mainly include camera-based image recognition algorithms, sensor-based and wifi signal-based detection algorithms.

The camera-based image recognition algorithm can usually realize a high-precision behavior detection system, but such a system requires people to be within the monitoring range of the camera, and it is easy to produce monitoring blind spots under the influence of obstacles and light. People's private life caused great interference.

Sensor-based detection algorithms require users to bind special devices to their bodies for behavior perception. Although such a system will not affect user privacy, it cannot satisfy the convenience of use.

The detection algorithm based on wifi signal recognizes body behavior by analyzing the different change patterns of human behavior on signal frequency, amplitude and other characteristics, because these signal characteristics contain less information, and the existing system does not have the ability to analyze complex body behavior , so that the system recognition accuracy and practicability are subject to certain limitations.

Contents of the invention

In order to solve the problems in the prior art, the present invention proposes a wireless non-contact method that recognizes human behavior by analyzing different change patterns of channel state information, and can achieve high recognition accuracy while satisfying convenience and security. Binding human behavior detection algorithm.

In order to achieve the above object, the technical solution adopted in the present invention is: comprise the following steps:

1) System deployment and model initialization: The wireless transmitter establishes a wifi field with stable waveforms, executes each meta-action in the wifi field, and the wireless receiver receives the wifi signal and calculates the CSI value of each meta-action, and calculates the CSI value of each meta-action After noise filtering, draw the change waveform diagram of the CSI phase angle of each meta-action according to the time sequence, save the waveform change sequence feature as the meta-action template sequence feature, and complete the initialization;

2) Feature extraction of the action sequence to be recognized: the user performs the action to be recognized in the wifi field, the wireless receiving end calculates and completes the CSI value of the action to be recognized, and after noise filtering the CSI value of the action to be recognized, draws the CSI phase of the action to be recognized The waveform diagram of the timing change of the angle, and save the waveform change feature as the action sequence feature to be recognized;

3) Behavior recognition:

3.1) Edge detection is performed on the time-series change waveform diagram of the CSI phase angle of the action to be recognized, and the start and end moments of the action execution are determined;

3.2) Use dynamic time warping to calculate the distance values between all meta-action template sequence features, and calculate the average value of the above distance values as the threshold T;

3.3) Use dynamic time warping to match the sequence features of the action sequence to be recognized with the sequence features of each meta-action template one by one, and calculate the distance value between the sequence feature of the action sequence to be recognized and the sequence features of each meta-action template in turn, if the distance values are not greater than the threshold T , then select the meta-action with the smallest distance value as the recognition result of the action to be recognized; if the distance values are greater than the threshold T, the recognition fails, and jump to step 2) to re-identify, if the distance values are still greater than the threshold T, then the The action sequence feature to be recognized is saved as a meta-action template sequence feature.

When the system is deployed in step 1), it is necessary to debug the wifi field in a static environment: the signal waveform transmitted by the wireless transmitter is X, the signal waveform received by the wireless receiver is Y, and the ratio of Y to X is static The CSI value in the environment is a complex number. The wireless receiving end draws the CSI phase angle timing waveform diagram. If the waveform fluctuation range is less than 0.2dBm, the waveform is considered stable; otherwise, after increasing the signal transmission power, draw the CSI phase angle timing waveform diagram. Until the waveform fluctuation range is less than 0.2dBm.

The CSI value of each meta-action in the step 1) and the CSI value of the action to be recognized in the step 2) are subjected to wavelet transform to realize noise filtering.

The wavelet transform uses the Dobesy wavelet db3 to perform noise filtering processing on the CSI value.

In the step 1), the wireless transmitter is a wireless router for wifi signal transmission, and the wireless receiver is a wireless router for wifi signal reception.

Compared with the prior art, the present invention uses the wireless transmitter to establish a wifi field. When the user walks or makes certain actions in the wifi coverage area, it will have a specific impact on the wifi channel, and the wireless receiver is used to receive the wifi signal and Calculate the CSI value of the action, extract the channel change characteristics, use the different change characteristics of the wifi channel when people move in the wifi field or perform certain actions, and analyze the different change patterns of the channel state information, using feature extraction and classification matching algorithms Carry out behavior recognition, combine human behavior and different change patterns of channels, so as to realize the recognition of human behavior with wifi channel characteristics, the present invention can satisfy convenience and safety while realizing high recognition accuracy, and does not need user Carrying any special equipment will not record the user's private life, which is convenient and easy to deploy, and has the characteristics of high security.

Further, when the system is deployed, it is necessary to debug the wifi field in a static environment, so that the wireless transmitter can transmit a stable wifi field, and the waveform fluctuation range of the CSI phase angle timing waveform diagram drawn by the wireless receiver is less than 0.2dBm, which is conducive to improving Accuracy of motion recognition.

Furthermore, the Dobessy wavelet db3 is selected to process the data. The db3 wavelet has the following two advantages: 1) It has good orthogonal symmetry, which is convenient for calculation and signal reconstruction; 2) The db3 wavelet can generate as many zero wavelets as possible Coefficients, which are good for data compression and noise removal. The third-order wavelet transform is the best way to obtain the signal after denoising, which not only effectively removes the noise, but also retains the local characteristics of the signal.

Description of drawings

Figure 1 is a flowchart of system deployment and model initialization;

Fig. 2 is a flow chart of feature extraction of an action sequence to be recognized;

Fig. 3 is a flow chart of behavior recognition;

Fig. 4 is a second-order discrete wavelet transform process diagram;

Figure 5a is a noisy feature sequence diagram obtained after two users perform the same action, and Figure 5b is a denoising feature sequence result diagram after third-order discrete wavelet transform;

Fig. 6 is a diagram of the dynamic time warping result of the feature sequence of the action to be recognized and the template meta-action.

Detailed ways

The present invention will be further explained below in conjunction with specific embodiments and accompanying drawings.

The present invention utilizes the different changing features generated by the wifi channel when the user moves in the wifi field or performs certain actions, and uses feature extraction and classification matching algorithms to perform behavior recognition, specifically including the following steps:

Step 1, see Figure 1, system deployment and model initialization:

1. Deploy two wireless routers indoors, one for wifi signal transmission and one for wifi signal reception;

2. Turn on two wireless routers at the same time, one transmits signal waveform X, and the other receives signal waveform Y, then the ratio of Y to X is the CSI in a static environment, and the value is a complex number;

3. Draw the CSI phase angle timing waveform diagram. If the waveform fluctuation range is less than 0.2dBm, the waveform is considered stable and skips to 5;

4. Increase the signal transmission power, skip to 2;

5. The user completes the specified meta-actions (such as waving, walking, kicking, squatting, etc.) in sequence indoors;

6. Calculate the CSI value of the channel when performing each action at the receiving end;

7. Perform noise filtering on the original CSI results;

8. Draw the change waveform diagram of the CSI phase angle according to the time sequence;

9. Save the characteristic sequence of the waveform as the characteristic pattern of the corresponding meta-action;

10. The initialization setting is completed, and the system returns;

Step 2, see Figure 2, feature extraction of action sequences to be recognized:

1. Turn on two wireless routers, and the transmitter sends signal waveform X;

2. The user performs an action to be recognized indoors;

3. The signal waveform received at the receiving end is Y, and the ratio of Y to X is used as the CSI value;

4. Perform noise filtering on the original CSI value;

5. Draw the waveform diagram of the timing change of the CSI phase angle;

6. The system returns;

Step 3, see Figure 3, behavior recognition:

1. Perform edge detection on the CSI phase waveform diagram to determine the start and end moments of action execution;

2. Use Dynamic Time Warping (DTW) to calculate the distance value between any two element action feature sequences, and calculate the average value of the above distance value as the threshold T;

3. Use DTW to match the action feature to be tested with the feature sequence of the meta-action template, and calculate the distance between the two in turn;

4. If the distance values are greater than the threshold T, jump to 6;

5. Select the meta-action with the smallest distance value as the recognition result of the action to be tested, and jump to 7;

6. The screen prompts that there is a recognition error, and jump to step 2—extraction of sequence features of the behavior to be recognized, re-execute the action and extract the sequence features;

7. The screen outputs the recognition result of the action to be tested;

8. The system returns.

The core method among the present invention is as follows:

1. Noise filtering is performed on CSI values in system deployment, model initialization and feature extraction of action sequences to be recognized, which is waveform noise filtering based on discrete wavelet transform:

The original CSI value contains a lot of noise information. The noise is mainly composed of Gaussian white noise in the environment and thermal noise of communication equipment. In order to improve the accuracy of the third step of behavior recognition, it is necessary to filter the noise in the signal. The reason why we cannot simply use a low-pass or high-pass filter for noise filtering is that we cannot determine the frequency range of the useful signal in advance and design a better filter. We use discrete wavelet transform to achieve noise filtering. Discrete wavelet transform decomposes and reconstructs the original noisy signal multiple times, and performs fine-grained multi-scale analysis, so as to achieve the best filtering effect. At the same time, the discrete wavelet transform can provide the optimal resolution in the time-frequency domain, and can more flexibly adapt to the impact of environmental changes on the data.

Selection of wavelet function: Observing that the original signal collected by us is relatively flat overall, we choose the Dobessy wavelet db3 to process the data. The db3 wavelet has the following two advantages: 1) It has good orthogonal symmetry, which is convenient for calculation and signal reconstruction; 2) The db3 wavelet can generate as many zero wavelet coefficients as possible, which is beneficial to data compression and noise elimination.

Discrete wavelet transform denoising principle: the output result of wavelet transform on a discrete sequence is a set of wavelet coefficients. These coefficients correspond to the high/low frequency components of the input sequence at different frequency scales, respectively. Set the high-frequency wavelet coefficients corresponding to the noise to zero, and reconstruct the signal with the changed set of wavelet coefficients to obtain the denoised signal. The difficulty of this process is to choose the appropriate order of wavelet transform, so as to separate the signal and noise under the optimal frequency resolution. After experimental verification, we choose the third-order wavelet transform to obtain the best signal effect after denoising, which not only effectively removes the noise, but also retains the local characteristics of the signal.

Referring to Fig. 4, it shows the second-order discrete wavelet transform process. Expressing the first-order wavelet transform with the formula is as follows:

Where x[2n-k] represents the original input signal; n represents the array index; k represents the loop summation variable, traversing from 0 to positive infinity; g[k] and h[k] represent low-pass and high-pass weights, respectively These coefficients are determined by the db3 wavelet; through the three functions wavedec(), appcoef(), and detcoef() in the Matlab wavelet transform tool set, the low-frequency wavelet coefficient x _{1, L} [n] and high-frequency wavelet coefficient x can be directly calculated _{1, H} [n], the subscript 1 represents the first-order wavelet transform, and these coefficients correspond to the high-frequency component and low-frequency component of the original signal, respectively. The second-order wavelet transform takes x1 _{, L} [n] as the input signal, and the formula for obtaining the second-order wavelet transform is as follows:

Thus, the second-order low-frequency wavelet coefficient x _{2, L} [n] and the second-order high-frequency wavelet coefficient x _{2, H} [n] are obtained;

The third-order wavelet transform takes x2 _{, L} [n] as the input signal, and the formula for obtaining the third-order wavelet transform is as follows:

Thus, the third-order low-frequency wavelet coefficient x _{3, L} [n] and the third-order high-frequency wavelet coefficient x _{3, H} [n] are obtained;

This is done in terms of x _1,L [n], x _1,H [n], x _2,L [n], x _2,H [n], x _3,L [n] and x _3,H [n] For signal analysis at different frequency resolutions, the formula for reconstructing the signal is as follows:

c _q [n] = ∑ _k g [n-2k] x _{q + 1, L} [n] ten ∑ _k h [n-2k] x _{q + 1, H} [n]

Where q represents the different orders of the wavelet transform, the signal of each order is reconstructed from the back to the front, and the finally obtained c ₁ [n] is the signal after denoising, and the noise filtering is completed. Figure 5a shows the noisy feature sequence graph obtained after two users perform the same action, and Figure 5b shows the denoising feature sequence result graph after the third-order discrete wavelet transform. The main steps of the denoising algorithm are as follows:

(1) Perform third-order discrete wavelet transform on the original noise-containing signal sequence of 192 points each time to obtain three layers of high-frequency and low-frequency wavelet coefficients;

(2) Set the high-frequency wavelet coefficients of the three layers to zero respectively;

(3) Using the modified wavelet coefficients to reconstruct the signal, the denoised signal sequence can be obtained.

2. Sequence feature matching

In this method, the degree of similarity between two sequences (usually represented by the distance between the two sequences) is calculated to determine the most likely action category of the action to be tested. Since different users perform actions at different times and at different speeds, the length of the feature sequence is different. That is to say, the characteristics between the two sequences are similar, but there is a possibility of misalignment in time, so one of the sequences needs to be warped under the time axis to achieve a better alignment effect. This method uses Dynamic Time Warping (Dynamic Time Warping, DTW) to achieve this purpose. The sequence feature matching algorithm is as follows:

(1) To calculate the distance between the action waveform sequence to be tested and the meta-action waveform sequence, first apply two n- _moment ×m matrices D and d, which are the cumulative distance matrix and the frame matching distance matrix respectively, where n _moments and m are respectively is the length of the action waveform sequence to be tested and the meta-action waveform sequence;

(2) Calculate the frame matching distance matrix between the two sequences by looping. The (i, j) element in the matrix represents the distance between the i-th element in the action waveform sequence to be tested and the j-th element in the meta-action waveform sequence ;

(3) Calculate the cumulative distance matrix D, let D(0,0)=0, for each point (i,j), calculate D(i,j)=d(i,j)+min(D(i -1, j), D(i-1, j-1), D(i, j-1)), where i, j respectively represent the i-th element in the action waveform sequence to be tested and the i-th element in the meta-action waveform sequence j elements;

(4) Choose D(n _moment , m) to represent the distance between the action waveform sequence to be tested and the meta-action waveform sequence.

Figure 6 shows the results of aligning the two sequences, and the solid and dotted lines represent the feature sequences of the action to be tested and the template action, respectively.

The present invention has the following advantages: compared with the prior art, using the human behavior detection technology in the present invention does not require the user to carry any special equipment, does not record the user's private life, and the recognition accuracy of the system is affected by obstacles in the static environment and Light is less affected. On the other hand, using dynamic time warping for feature matching can adapt to the behavior detection of different users, and the system can meet the convenience and security while achieving high recognition accuracy.

Claims (4)

1. a kind of wireless unbundling human body behavioral value algorithm, which comprises the following steps:

1) system deployment and model initialization: wireless transmitting terminals establish the field wifi of waveform stabilization, execute in wifi each Metaaction, wireless receiving end receive wifi signal and calculate the CSI value of each metaaction, and to the CSI value of each metaaction into After Row noise filtering, the variation waveform diagram at the phase angle CSI of each metaaction is drawn according to timing, and it is special to save waveform change sequence Sign is used as metaaction template sequence feature, completes initialization；

2) action sequence feature extraction to be identified: user executes movement to be identified in wifi, and wireless receiving end, which calculates, completes The CSI value of movement to be identified, and treat identification maneuver CSI value carry out noise filtering after, draw the CSI phase of movement to be identified The timing variations waveform diagram at angle saves waveform variation characteristic as action sequence feature to be identified；

3) Activity recognition:

3.1) the timing variations waveform diagram for treating the phase angle CSI of identification maneuver carries out edge detection, determines that movement executes Beginning and end time；

3.2) calculate the distance value of all metaaction template sequence features between any two using dynamic time warping, and to above away from It averages from value as threshold value T；

3.3) identification maneuver sequence signature is treated using dynamic time warping to carry out one by one with each metaaction template sequence feature Matching, and the distance value between action sequence feature to be identified and each metaaction template sequence feature is successively calculated, if distance value It is not to be all larger than threshold value T, then selects recognition result of the smallest metaaction of distance value as movement to be identified；If distance value is big In threshold value T, then recognition failures, go to step and 2) re-start identification, if distance value is still all larger than threshold value T, this is waited for Identification maneuver sequence signature is saved as metaaction template sequence feature；

It needs to debug wifi under static situation when system deployment in the step 1): wireless transmitting terminals transmitting Signal waveform is X, and the signal waveform that wireless receiving termination receives is that the ratio of Y, Y and X are the CSI value under static situation, value For plural number, the phase angle CSI timing waveform is drawn at wireless receiving end, if waveform fluctuation range is less than 0.2dBm, then it is assumed that waveform Stablize；Otherwise after improving signal transmission power, then the phase angle CSI timing waveform is drawn, until waveform fluctuation range is less than 0.2dBm；

Distance value calculating process in the step 3.3) between motion characteristic to be identified and each metaaction template characteristic is as follows:

(1) two n are set_SquareThe matrix D and d of × m, respectively Cumulative Distance matrix D and frame matching distance matrix d, wherein n_SquareWith m The length of action waveforms sequence and metaaction wave sequence respectively to be identified；

(2) pass through the frame matching distance matrix d between cycle calculations action waveforms sequence to be identified and metaaction wave sequence, square In battle array in (i, j) element representation action waveforms sequence to be identified in i-th of element and metaaction wave sequence between j-th of element Distance；

(3) Cumulative Distance matrix D is calculated, D (0,0)=0 is enabled, each point (i, j) is calculated separately:

D (i, j)=d (i, j)+min (D (i-1, j), D (i-1, j-1), D (i, j-1)), wherein i, j respectively indicate to be identified dynamic Make i-th of element and j-th of element in metaaction wave sequence in wave sequence；

(4) D (n is selected_Square, m) and indicate the distance between action waveforms sequence to be identified and metaaction wave sequence.

2. a kind of wireless unbundling human body behavioral value algorithm according to claim 1, which is characterized in that the step 1) it is made an uproar in the CSI value for treating identification maneuver in the CSI value of each metaaction and the step 2) using wavelet transformation to realize Sound filtering.

3. a kind of wireless unbundling human body behavioral value algorithm according to claim 2, which is characterized in that the small echo Transformation carries out noise filtering processing to CSI value using the western small echo db3 of more shellfishes.

4. a kind of wireless unbundling human body behavioral value algorithm according to claim 1, which is characterized in that the step 1) wireless transmitting terminals are the wireless router emitted for wifi signal in, and wireless receiving end is for the received nothing of wifi signal Line router.