CN111626174A - Attitude robust motion recognition method based on channel state information - Google Patents

Abstract

The invention belongs to the technical field of single-user gesture robustness, and in particular relates to a gesture robust action recognition method based on channel state information. Intelligent life is the goal pursued by people in the future, and the accurate realization of human-computer interaction is an urgent problem to be solved at present. Existing action recognition methods based on channel state information have low accuracy in recognizing actions of the same user at different times. The invention improves the shortcomings of the existing methods by introducing a method for extracting stable features, extracts stable features with little change, and improves the recognition accuracy.

Description

A robust gesture recognition method based on channel state information

technical field

The invention belongs to the technical field of single-user gesture robustness, and in particular relates to a gesture robust action recognition method based on channel state information.

Background technique

Nowadays, there are not many studies on the actions of the same user at different times in the research on CSI action recognition. However, considering the practicality of the system, it is also necessary to study the actions of the same user. Since even the same user performs the same action, the magnitude or direction of the action will be different, which will cause the easily affected channel state information to change, that is, the fingerprint information to change. It will affect the final recognition of the action. This paper will propose an improved method for the above problems. The idea proposed in this paper is based on the similarities of signature verification. In signature verification, since each person's signature will have some features that are different from others and remain unchanged, this feature fragment is extracted as the main part of identification. Similarly, since each person has different living habits, even if he makes the same action at different times, the channel state information will change due to different angles or speeds, but in the action sequence he does, there will be A major part is one that does not change or even changes very little. So rely on this idea to come up with a solution.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a gesture robust action recognition method based on channel state information.

The object of the present invention is achieved through the following technical solutions: comprise the following steps:

Step 1: Arrange the transmitter and receiver, and the receiver obtains the channel state information;

Step 2: Calculate the amplitude variance of each subcarrier in the channel state information, and extract the N key subcarriers with the largest variance;

Step 3: Use the principal component analysis method to reduce the dimension and denoise the extracted key sub-carriers;

Step 4: Further denoising using discrete wavelet transform;

Step 5: Extract stable features;

Step 5.1: Divide the subcarriers into L parts, and the waves to be matched are called C and D;

Step 5.2: Set the threshold T, let i=1;

Step 5.3: judge whether i is less than L+1; if i is greater than L+1, end the extraction of stable features; if i is less than L+1, execute step 5.4;

Step 5.4: Calculate C _i and D _i ;

C _i =X _i+1 -X _i

D _i =Y _i+1 -Y _i

Among them, X _i represents the ordinate of the ith point in C, that is, the amplitude; Y _i represents the amplitude of the ith point in D;

Step 5.5: judge whether C _i and D _i have the same sign; if C _i and D _i have different signs, delete C _i and D _i , set i=i+1, and return to step 5.3; if C _i and D _i have the same sign, Then go to step 5.6;

Step 5.6: judge whether C _i -D _i is less than T; if C _i -D _i is less than T, keep C _i and D _i , set i=i+1, and return to step 5.3; if C _i -D _i is greater than or equal to T, then delete C _i and D _i , set i=i+1, and return to step 5.3;

Step 6: Use the random forest method to classify the extracted stable features to obtain pose robust action recognition results.

The beneficial effects of the present invention are:

Intelligent life is the goal pursued by people in the future, and the accurate realization of human-computer interaction is an urgent problem to be solved at present. Existing action recognition methods based on channel state information have low accuracy in recognizing actions of the same user at different times. The invention improves the deficiencies of the existing methods by introducing a method for extracting stable features, extracts stable features with little change, and improves the recognition accuracy.

Description of drawings

FIG. 1 is a frame diagram of the present invention.

FIG. 2 is a flow chart of stable feature extraction in the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

As shown in FIG. 1 , the method includes four modules: CSI data acquisition, signal preprocessing, feature extraction, and behavior recognition.

1) Place the transmitter and receiver at any position in the room, and the user stands between the transmitter and the receiver to perform actions to obtain channel status information;

2) Extracting a plurality of sensitive key sub-carriers from the received original information;

3) Use PCA dimensionality reduction and denoising on the key subcarriers to obtain the subcarriers that can best reflect the action information;

4) Further use wavelet transform to denoise;

5) Extract stable features;

6) Use random forest method to classify actions;

1) FIG. 1 is a framework diagram of the present invention. First, the original CSI data of the wireless signal is collected from the receiver, and then data preprocessing, feature extraction, and action recognition are performed in sequence.

2) In the data collection stage, place the transmitter and receiver at any position in the room, but it is better to separate them by a certain distance. The user stands between the transmitter and the receiver to perform actions. Root antenna TP-Link802.11n wireless router, the receiver is a levono laptop with Intel 5300 network card and three external antennas, the sampling frequency is set to 1000Hz, and the receiver is used to obtain channel status information.

3) Since the OFDM technology is used to divide the channel into 30 sub-carriers, but each sub-carrier has different sensitivity to the channel, so the sensitivity is first quantized, and the amplitude variance of each sub-carrier is calculated. Sensitivity, that is, the larger the variance is, the more critical subcarriers are needed.

4) However, the selected sub-carriers still contain a lot of noise and even some of them are not related to the action itself. Therefore, further selection is required to find out the sub-carriers that can best reflect the action information. Principal component analysis (PCA) can be used in this step.

5) After processing, although a part of the noise has been filtered, there is noise caused by the equipment itself and other problems, so further denoising is required. This part of the noise is removed using discrete wavelet transform (DWT).

6) Use Figure 2 to extract stable features.

7) Use the method of random forest to classify the results.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (1)

1. A posture robust motion recognition method based on channel state information is characterized by comprising the following steps:

step 1: arranging a transmitter and a receiver, wherein the receiver acquires channel state information;

step 2: calculating the amplitude variance of each subcarrier in the channel state information, and extracting N key subcarriers with the largest variance;

and step 3: reducing and denoising the extracted key subcarriers by using a principal component analysis method;

and 4, step 4: further denoising using discrete wavelet transform;

and 5: extracting stable characteristics;

step 5.1: dividing the sub-carrier into L parts, and calling waves to be matched as C and D;

step 5.2: setting a threshold value T, and enabling i to be 1;

step 5.3: judging whether i is smaller than L + 1; if i is larger than L +1, finishing the extraction of the stable features; if i is less than L +1, executing step 5.4;

step 5.4: calculating C_iAnd D_i；

C_i＝X_i+1-X_i

D_i＝Y_i+1-Y_i

Wherein, X_iRepresents the ordinate, namely the amplitude, of the ith point in C; y is_iRepresenting the amplitude of the ith point in D;

step 5.5: judgment C_iAnd D_iWhether the numbers are the same; if C_iAnd D_iIf the numbers are different, C is deleted_iAnd D_iMaking i equal to i +1, and returning to the step 5.3; if C_iAnd D_iIf yes, executing step 5.6;

step 5.6: judgment C_i-D_iWhether less than T; if C_i-D_iIf less than T, then C is retained_iAnd D_iMaking i equal to i +1, and returning to the step 5.3; if C_i-D_iIf T is greater than or equal to T, C is deleted_iAnd D_iMaking i equal to i +1, and returning to the step 5.3;

step 6: and classifying the extracted stable features by using a random forest method to obtain a posture robust motion recognition result.

Images