An online wifi-aware low-delay activity signal segmentation method and device

Abstract

The application provides an online wifi-aware low-delay activity signal segmentation method and device, and the steps of the method comprise: collecting channel state information of a wifi signal to perform windowing, to obtain a channel state information window sequence; performing dimension reduction on the channel state information window sequence, processing the dimension-reduced channel state information window sequence based on a moving variance algorithm, to obtain a moving variance sequence; sequentially extracting each data point in the moving variance sequence according to a time parameter, if the variance values of the first preset number of data points after a certain data point are all less than the product of the variance value of the data point and a preset ratio parameter, and there is no data point before the data point with a variance value greater than the data point, then taking the variance value of the data point as a calibration variance value; calculating a granularity awareness threshold based on the calibration variance value, calculating a front-end segmentation point and a rear-end segmentation point based on the granularity awareness threshold, and segmenting the channel state information sequence based on the front-end segmentation point and the rear-end segmentation point.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wireless sensing technology, and in particular to a low-delay activity signal segmentation method and device for online WiFi sensing. BACKGROUND

[0002] In recent years, wireless sensing technology has attracted more and more attention, which uses WiFi, Bluetooth and other wireless radio frequency signals to sense the behavior of people or other objects, and then realizes intelligent applications. The accuracy, efficiency and practicality of the sensing system determine the quality of user experience. WiFi signals are ubiquitous and are not affected by line-of-sight and lighting, so WiFi-based wireless sensing has attracted extensive research. Compared with RSSI (Received Signal Strength Indicator), CSI (Channel State Information) contains more fine-grained and stable information. It is resistant to background noise and sensitive to human activity. Therefore, CSI is usually used for WiFi sensing applications. CSI contains amplitude and phase information. The CSI phase is often disturbed by uncertain factors such as carrier frequency offset and sampling frequency offset. While the amplitude of the CSI is usually a reliable and easy-to-read measure for activity extraction and classification, so the CSI amplitude is a common feature of WiFi sensing.

[0003] There are two main stages of WiFi-based activity recognition: activity extraction and activity classification. Activity extraction is achieved by segmenting the signal segments containing human activity in the CSI signal. While correct activity signal segmentation is crucial for the accuracy of subsequent activity classification.

[0004] The existing threshold-based signal segmentation method sets the threshold according to the granularity of the activity it studies. In actual scenarios, coarse-grained activities and fine-grained activities usually occur randomly in alternation, and the activity recognition system needs to continuously segment coarse-grained activity (such as walking, squatting, etc.) signals and fine-grained activity (such as hand gestures) signals. The existing threshold-based method only sets appropriate thresholds for activities with a certain granularity in a specific application scenario. However, it is difficult to set thresholds suitable for all types of activities before extracting any activity, as the granularity of the activity cannot be predicted in advance. For existing methods, the threshold determined for the signal of one granularity may not be suitable for the segmentation of the signal of another granularity, and it is difficult to flexibly set the threshold for signal segmentation, resulting in a decrease in segmentation accuracy and ultimately affecting the accuracy of activity recognition. SUMMARY

[0005] In view of this, embodiments of the present application provide a low-delay activity signal segmentation method for online WiFi sensing to eliminate or improve one or more defects in the prior art.

[0006] One aspect of the present application provides an online wifi-aware low-delay activity signal segmentation method, the steps of the method comprising:

[0007] Channel state information of the wifi signal is collected by a network card, a plurality of data points in the channel state information are constructed into a channel state information sequence, and the channel state information sequence is windowed according to time sequence in a sliding window manner to obtain a channel state information window sequence;

[0008] The channel state information window sequence is reduced in dimension based on a principal component analysis method, and the channel state information window sequence after dimension reduction is processed based on a moving variance algorithm to obtain a moving variance sequence;

[0009] Each data point in the moving variance sequence is extracted according to time parameter sequence, if variance values of a first preset number of data points after a certain data point are all less than a product of the variance value of the data point and a preset ratio parameter, and there is no data point with a variance value greater than the variance value of the data point before the data point, the variance value of the data point is taken as a calibrated variance value;

[0010] A granularity awareness threshold is calculated based on the calibrated variance value, a front-end segmentation point and a rear-end segmentation point are calculated in the moving variance sequence based on the granularity awareness threshold, and the channel state information sequence is segmented based on the front-end segmentation point and the rear-end segmentation point.

[0011] In the above scheme, the channel state information sequence is first processed by the moving variance algorithm, since the channel state information during the active part is greater than the channel state information during the static part, the moving variance algorithm can capture these differences, then the granularity awareness threshold is calculated based on the calibrated variance value in the moving variance sequence, and the front-end segmentation point and the rear-end segmentation point are further calculated, and the channel state information sequence is segmented based on the time of the front-end segmentation point and the rear-end segmentation point, on the one hand, the granularity awareness threshold can be flexibly set, on the other hand, the granularity awareness threshold in the present application is not directly applied to the channel state information sequence, but the front-end segmentation point and the rear-end segmentation point are calculated according to the granularity awareness threshold, and then the channel state information sequence is segmented, which improves the segmentation accuracy and ensures the accuracy of subsequent activity recognition of the segmented data.

[0012] In some embodiments of the present application, in the step of calculating the granularity awareness threshold based on the calibrated variance value, the granularity awareness threshold is calculated according to a preset ratio parameter by using the following formula:

[0013] Granularity awareness threshold = calibrated variance value * ratio parameter.

[0014] In some embodiments of the present invention, the step of calculating the front-end segmentation point and the back-end segmentation point in the moving variance sequence based on the granularity-aware threshold includes:

[0015] Based on the time parameters corresponding to each data point in the moving variance sequence, the first data point whose variance value is less than the granularity perception threshold is matched according to the time parameters and the order of the data points corresponding to the calibration variance values. The matched data points are used as back-end segmentation points.

[0016] Based on the time parameters corresponding to each data point in the moving variance sequence, the first data point with a variance value greater than the granularity perception threshold is matched according to the time parameter in the order of the starting data points, and the matched data point is used as the front-end segmentation point.

[0017] In some embodiments of the present invention, in the step of sequentially matching the first data point with a variance value greater than the granularity perception threshold according to the time parameter and using the matched data point as the front-end segmentation point, a second preset number of data points are collected after the data point with the first variance value greater than the granularity perception threshold. If the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points is greater than the segmentation threshold, then the data point with the first variance value greater than the granularity perception threshold is used as the front-end segmentation point. If the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points is not greater than the segmentation threshold, then the next data point with a variance value greater than the granularity perception threshold is matched after the data point with the first variance value greater than the granularity perception threshold, and the above steps are repeated until the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points after the next data point with a variance value greater than the granularity perception threshold is greater than the segmentation threshold, and then that data point is used as the front-end segmentation point.

[0018] In some embodiments of the present invention, in the step of sequentially matching the data points corresponding to the calibrated variance values ​​according to the time parameter, and using the first data point with a variance value less than the granularity perception threshold as the back-end segmentation point, a third preset number of data points are collected after the first data point with a variance value less than the granularity perception threshold. If the proportion of data points with a variance value less than the granularity perception threshold among the third preset number of data points is greater than the segmentation threshold, then the first data point with a variance value less than the granularity perception threshold is used as the back-end segmentation point. If the proportion of data points with a variance value less than the granularity perception threshold among the third preset number of data points is greater than the segmentation threshold, then the next data point with a variance value less than the granularity perception threshold is matched after the first data point with a variance value less than the granularity perception threshold, and the above steps are repeated until the proportion of data points with a variance value less than the granularity perception threshold among the third preset number of data points after the next data point with a variance value less than the granularity perception threshold is greater than the segmentation threshold, and then that data point is used as the back-end segmentation point.

[0019] In some embodiments of the present invention, the third predetermined number is greater than the second predetermined number.

[0020] The segmentation threshold is a percentage value.

[0021] In some embodiments of the present invention, in the step of segmenting the channel state information sequence based on the front-end segmentation point and the back-end segmentation point, the time parameters corresponding to the front-end segmentation point and the back-end segmentation point are obtained, and the data points between the time parameters of the front-end segmentation point and the time parameters of the back-end segmentation point are segmented from the channel state information sequence.

[0022] In some embodiments of the present invention, the steps before performing dimensionality reduction on the channel state information window sequence based on principal component analysis further include preprocessing the channel state information window sequence, the preprocessing steps including amplitude limiting and noise reduction.

[0023] The amplitude limiting process involves deleting data points in the channel state information window sequence whose amplitude is greater than the preset amplitude threshold, based on the preset amplitude threshold.

[0024] The denoising process involves using discrete wavelet transform to denoise the channel state information window sequence.

[0025] In some embodiments of the present invention, the preprocessing step includes compression processing, wherein the compression processing step is to compress the channel state information window sequence using an equidistant sampling method.

[0026] In some embodiments of the present invention, the step of processing the dimension-reduced channel state information window sequence based on the moving variance algorithm to obtain the moving variance sequence includes,

[0027] Obtain the preset sub-window size of the moving variance algorithm. In the channel state information window, slide the sub-window starting from the data point at the beginning time of the channel state information window sequence according to the time order. Calculate the variance of the amplitude of the data point in the sub-window after each slide. Use the calculated variance as the variance value of the earliest time point in the sub-window in the moving variance sequence.

[0028] In some embodiments of the present invention, the step of dividing the channel state information sequence into windows according to time order using a sliding window method to obtain a channel state information window sequence further includes:

[0029] The data points collected in this sliding window event are obtained using a sliding window method.

[0030] Obtain the data points after the time point of the back-end segmentation point in the previous channel state information window sequence;

[0031] The data points collected in this sliding window are merged with the data points after the time point of the back-end segmentation point in the previous channel state information window sequence to obtain the current channel state information window sequence.

[0032] In some embodiments of the present invention, the method further includes the following steps:

[0033] Obtain the segmented sequence obtained by segmenting the channel state information sequence based on the front-end segmentation point and the back-end segmentation point, and input the parameters in the segmented sequence into a preset classifier;

[0034] The action type is determined based on the output parameters of the classifier.

[0035] The present invention also provides a low-latency activity signal segmentation device for online Wi-Fi sensing. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0036] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0037] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0038] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0039] Figure 1 This is a schematic diagram of one embodiment of the low-latency activity signal segmentation method for online Wi-Fi sensing according to the present invention;

[0040] Figure 2 This is a schematic diagram of another embodiment of the low-latency activity signal segmentation method for online Wi-Fi sensing according to the present invention;

[0041] Figure 3 This is a schematic diagram illustrating the implementation process of the method of the present invention;

[0042] Figure 4 A schematic diagram of a sequence of channel state information windows represented as an image;

[0043] Figure 5 A schematic diagram showing the channel state information window sequence after dimensionality reduction using principal component analysis as an image;

[0044] Figure 6 A schematic diagram of the moving variance sequence represented as an image;

[0045] Figure 7 A schematic diagram of a channel state information sequence without removing anomalous peaks, represented as an image;

[0046] Figure 8 A schematic diagram of a channel state information sequence with abnormal peaks removed, represented as an image. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0048] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0049] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0050] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0051] Existing technologies also include using deep learning models to perceive activity granularity, enabling perception applicable to activities of different granularities and eliminating reliance on empirical threshold settings. However, deep learning methods incur significant overhead and cannot be applied to real-world scenarios requiring online activity signal segmentation, thus remaining confined to offline applications.

[0052] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0053] To solve the above problems, such as Figure 1 , 3 As shown, this invention proposes a low-latency activity signal segmentation method for online Wi-Fi sensing, the method comprising the following steps:

[0054] Step S100: Collect channel state information of Wi-Fi signals using a network card, construct a channel state information sequence from multiple data points in the channel state information, and divide the channel state information sequence into windows according to time order using a sliding window method to obtain a channel state information window sequence as shown below. Figure 4 As shown;

[0055] In some embodiments of the present invention, the network card can be an Intel 5300 network card, and the channel state information sequence is a sequence of data points of the channel state information of the Wi-Fi signal constructed in chronological order. In the channel state information sequence, the amplitude is recorded for each time parameter corresponding to each data point.

[0056] In some embodiments of the present invention, the channel state information sequence is divided into windows according to time order using a sliding window method to obtain a channel state information window sequence, and each channel state information window sequence obtained by sliding window is a part of the channel state information sequence.

[0057] To achieve online activity signal segmentation using the above scheme, we introduce a sliding window mechanism and ensure that all processing steps for activity signal segmentation are completed within the sliding window duration. Furthermore, this invention reserves sufficient time for the activity classification process. The lightweight algorithm meets the requirement of low processing time.

[0058] Step S200, as follows Figure 5 As shown, the channel state information window sequence is reduced in dimensionality based on principal component analysis, such as... Figure 6 As shown, the dimension-reduced channel state information window sequence is processed based on the moving variance algorithm to obtain the moving variance sequence;

[0059] In some embodiments of the present invention, principal component analysis is a multivariate statistical method that transforms multiple indicators into a few comprehensive indicators. The comprehensive indicators here are the principal components. Each principal component is a linear combination of the original variables, which are independent of each other and retain most of the information of the original variables. Its essence is to seek comprehensive substitutes for related variables through the correlation of the original variables, so as to ensure that the information loss during the transformation process is minimized while completing the data dimensionality reduction.

[0060] In some embodiments of the present invention, the data points in the channel state information window sequence can be 30-dimensional, 60-dimensional, or 90-dimensional, etc.

[0061] The moving variance sequence obtained by the above moving variance algorithm clearly shows the degree of data change. The degree of change can be changed by controlling the moving coefficient. The larger the moving coefficient, the smaller the degree of change may be. It mainly depends on the change of data. The moving variance algorithm can be the moving variance algorithm of MATLAB.

[0062] Each data point in the moving variance sequence has a corresponding time parameter with a variance value.

[0063] Step S300: Extract each data point in the moving variance sequence according to the time parameter order. If the variance values ​​of the first preset number of data points after a certain data point are all less than the product of the variance value of the data point and the preset ratio parameter, and there are no data points before the data point with a variance value greater than the variance value of the data point, then the variance value of the data point is taken as the calibration variance value.

[0064] By adopting the above scheme and under the above conditions: if the variance values ​​of the first preset number of data points after a certain data point are all less than the product of the variance value of the data point and the preset ratio parameter, and there are no data points before the data point with a variance value greater than the variance value of the data point, the maximum variance value in a continuous data segment can be accurately found, which is convenient for determining the calibration variance value in the data caused by an action, and can more accurately determine the granularity perception threshold in the data caused by an action.

[0065] Step S400: Calculate the granularity-aware threshold based on the calibration variance value; calculate the front-end segmentation point and the back-end segmentation point in the moving variance sequence based on the granularity-aware threshold; and segment the channel state information sequence based on the front-end segmentation point and the back-end segmentation point.

[0066] In some embodiments of the present invention, after each extraction of the channel state information window sequence, the front-end segmentation point and the back-end segmentation point are calculated based on the channel state information window sequence, and a segmentation is performed in the channel state information sequence.

[0067] The above scheme first uses a moving variance algorithm to process the channel state information sequence. Since the changes in channel state information during the active phase are greater than those during the static phase, the moving variance algorithm can easily capture these differences. Then, a granularity-aware threshold is calculated based on the calibration variance value in the moving variance sequence, and the front-end segmentation point and back-end segmentation point are further calculated. The channel state information sequence is segmented based on the time of the front-end segmentation point and the back-end segmentation point. This scheme can flexibly set the granularity-aware threshold. On the other hand, the granularity-aware threshold of this scheme does not directly apply to the channel state information sequence. Instead, the front-end segmentation point and the back-end segmentation point are calculated based on the granularity-aware threshold before segmenting the channel state information sequence, which improves the segmentation accuracy, ensures the accuracy of subsequent activity identification of the segmented data, and enables real-time processing, thus improving processing efficiency.

[0068] In the specific implementation process, after processing the channel state information window sequence of one window, the next channel state information window sequence obtained by the sliding window is processed, the front-end segmentation point and the back-end segmentation point are calculated, and the channel state information sequence is segmented based on the front-end segmentation point and the back-end segmentation point.

[0069] In some embodiments of the present invention, in the step of calculating the granularity perception threshold based on the calibration variance value, the granularity perception threshold is calculated using the following formula according to a preset ratio parameter:

[0070] Granularity perception threshold = calibration variance * ratio parameter.

[0071] In some embodiments of the present invention, the ratio parameter may be set to 4%.

[0072] In practical implementation, we found that for a single activity, whether coarse-grained or fine-grained, setting the threshold to 4% of the maximum value within the activity window allows for accurate segmentation. For offline CSI streams, it's easy to control that only one activity exists within the sliding window, or to obtain the approximate location of the target activity, thus easily ensuring that only one activity is within the threshold calculation range. However, for online CSI streams, since the timing of activity occurrences cannot be predicted, it's difficult to ensure that only one activity exists within the sliding window or to obtain the location of the target activity. If the calculation range is the entire sliding window, the threshold for segmenting the target activity signal might originate from another activity when the maximum variance of another activity is higher than the maximum variance of the target activity, leading to inaccurate signal segmentation. To address this issue, this solution proposes a threshold algorithm that iterates through and updates the maximum value online to obtain a granularity-aware threshold. Thanks to linear traversal, the time complexity of threshold calculation is low. The threshold is calculated based on a ratio parameter and a calibration variance value, thus giving it the ability to perceive activity granularity.

[0073] Using the above scheme, the corresponding granularity-aware threshold can be calculated based on the calibrated variance value in the moving variance sequence. It can be widely applied to coarse-grained or fine-grained activity signals, flexibly set the granularity-aware threshold, and improve the segmentation accuracy of moving variance sequences based on the granularity-aware threshold.

[0074] In some embodiments of the present invention, the step of calculating the front-end segmentation point and the back-end segmentation point in the moving variance sequence based on the granularity-aware threshold includes:

[0075] Based on the time parameters corresponding to each data point in the moving variance sequence, the first data point whose variance value is less than the granularity perception threshold is matched according to the time parameters and the order of the data points corresponding to the calibration variance values. The matched data points are used as back-end segmentation points.

[0076] Based on the time parameters corresponding to each data point in the moving variance sequence, the first data point with a variance value greater than the granularity perception threshold is matched according to the time parameter in the order of the starting data points, and the matched data point is used as the front-end segmentation point.

[0077] like Figure 7 As shown, in some embodiments of the present invention, in the step of calculating the front-end segmentation point and the back-end segmentation point, the first data point with a variance value greater than the granularity perception threshold after the starting data point can be directly used as the front-end segmentation point, and the first data point with a variance value less than the granularity perception threshold after the data point corresponding to the calibration variance value can be used as the back-end segmentation point.

[0078] like Figure 7 , 8 As shown, in some embodiments of the present invention, in the step of sequentially matching the first data point with a variance value greater than the granularity perception threshold according to the time parameter and using the matched data point as the front-end segmentation point, a second preset number of data points are collected after the data point with the first variance value greater than the granularity perception threshold. If the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points is greater than the segmentation threshold, then the data point with the first variance value greater than the granularity perception threshold is used as the front-end segmentation point. If the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points is not greater than the segmentation threshold, then the next data point with a variance value greater than the granularity perception threshold is matched after the data point with the first variance value greater than the granularity perception threshold, and the above steps are repeated until the proportion of data points with a variance value greater than the granularity perception threshold among the second preset number of data points after the next data point with a variance value greater than the granularity perception threshold is greater than the segmentation threshold, and then that data point is used as the front-end segmentation point.

[0079] If the above scheme is adopted and the first data point with a variance value greater than the granularity perception threshold is directly used as the front-end segmentation point, it is impossible to avoid abnormal peaks caused by interference information. This scheme can further calculate the second preset number of data points after the data point to avoid the influence of abnormal peaks on the front-end segmentation point and further improve the accuracy of the final judgment action.

[0080] like Figure 7 , 8 As shown, in some embodiments of the present invention, in the step of sequentially matching the data points corresponding to the calibrated variance values ​​according to the time parameter, and using the matched data points as back-end segmentation points, a third preset number of data points are collected after the data points with variance values ​​less than the granularity perception threshold. If the proportion of data points with variance values ​​less than the granularity perception threshold among the third preset number of data points is greater than the segmentation threshold, then the data point with the first variance value less than the granularity perception threshold is used as the back-end segmentation point. If the proportion of data points with variance values ​​less than the granularity perception threshold among the third preset number of data points is greater than the segmentation threshold, then the next data point with variance values ​​less than the granularity perception threshold is matched after the data point with the first variance value less than the granularity perception threshold, and the above steps are repeated until the proportion of data points with variance values ​​less than the granularity perception threshold among the third preset number of data points after the next data point with variance values ​​less than the granularity perception threshold is greater than the segmentation threshold, and then that data point is used as the back-end segmentation point.

[0081] Using the above scheme, if the first data point with a variance value less than the granularity perception threshold after the data point corresponding to the calibration variance value is directly used as the back-end segmentation point, it is easy to ignore the effective data points after the data point with a variance value less than the granularity perception threshold. This scheme further calculates the second or third predetermined number of data points after the data point. On the one hand, it can avoid ignoring the effective data points after the data point with a variance value less than the granularity perception threshold. On the other hand, it can avoid the abnormal peaks after the data point affecting the determination of the back-end segmentation point, and further improve the accuracy of the final judgment action.

[0082] In some embodiments of the present invention, if either the front-end segmentation point or the back-end segmentation point is not extracted, the extraction is determined to have failed.

[0083] The extraction failure scenarios include:

[0084] In the moving variance sequence, the proportion of data points with variance values ​​greater than the granularity perception threshold within the second preset number of data points after any data point whose variance value is greater than the granularity perception threshold is not greater than the segmentation threshold;

[0085] In the moving variance sequence, the proportion of data points with variance values ​​less than the granularity perception threshold within the third preset number of data points after any data point with a variance value less than the granularity perception threshold is not greater than the segmentation threshold.

[0086] In some embodiments of the present invention, in the step of segmenting the channel state information sequence based on the front-end segmentation point and the back-end segmentation point, the time parameters corresponding to the front-end segmentation point and the back-end segmentation point are obtained, and the data points between the time parameters of the front-end segmentation point and the time parameters of the back-end segmentation point are segmented from the channel state information sequence.

[0087] In some embodiments of the present invention, the steps before performing dimensionality reduction on the channel state information window sequence based on principal component analysis further include preprocessing the channel state information window sequence, the preprocessing steps including amplitude limiting and noise reduction.

[0088] The amplitude limiting process involves deleting data points in the channel state information window sequence whose amplitude is greater than the preset amplitude threshold, based on the preset amplitude threshold.

[0089] The denoising process involves using discrete wavelet transform to denoise the channel state information window sequence.

[0090] In some embodiments of the present invention, the preset amplitude threshold can be 35.

[0091] Using the above scheme, the Discrete Wavelet Transform (DWT) method can effectively preserve the peak values ​​and abrupt changes of the desired useful signal in the original signal. However, when using Fourier transform analysis for filtering, since the useful signal is concentrated in the low-frequency range while noise is concentrated in the high-frequency range, low-pass filtering cannot effectively distinguish between the high-frequency components of the desired signal and the high-frequency interference caused by noise. If the low-pass filter is too narrow, a portion of the desired signal will be filtered out as noise, erasing its morphological information and causing distortion of the original signal. After DWT noise filtering, the signal segmentation features are enhanced. Furthermore, since DWT has dedicated fast calculation functions in different programming languages, it is suitable for online activity recognition systems requiring low latency.

[0092] In some embodiments of the present invention, the preprocessing step includes compression processing, wherein the compression processing step is to compress the channel state information window sequence using an equidistant sampling method.

[0093] In some embodiments of the present invention, the equidistant sampling method can be the equidistant sampling method of MATLAB.

[0094] Using the above approach, this invention achieves a sufficiently low processing time (approximately 0.3 seconds per activity) on commonly used hardware such as laptops. However, in real-world scenarios, users may need to deploy the activity recognition system on resource-constrained hardware (such as embedded devices, Raspberry Pi, etc.), which will increase the processing time and impact the user experience. Reducing the sampling rate through data compression can significantly reduce processing time, but it will sacrifice segmentation accuracy to some extent. To balance the trade-off between efficiency and performance, users can set a specified compression ratio to compress data before other preprocessing steps. Compression is optional and usually does not need to be enabled; the compression method used is the most common isometric sampling method. Even though compression will reduce segmentation accuracy to some extent, this invention can still guarantee usable CSI segmentation accuracy even with a compression ratio of 20x.

[0095] In some embodiments of the present invention, the step of processing the dimension-reduced channel state information window sequence based on the moving variance algorithm to obtain the moving variance sequence includes,

[0096] Obtain the preset sub-window size of the moving variance algorithm. In the channel state information window, slide the sub-window starting from the data point at the beginning time of the channel state information window sequence according to the time order. Calculate the variance of the amplitude of the data point in the sub-window after each slide. Use the calculated variance as the variance value of the earliest time point in the sub-window in the moving variance sequence.

[0097] In some embodiments of the present invention, the sub-window, according to the time sequence, initially starts from the first data point in the channel state information window sequence, covering multiple data points. The variance value calculated for the first time is used as the variance value of the time point corresponding to the first data point in the channel state information window sequence in the moving variance sequence. The sub-window moves backward by one data point each time, and the variance of the amplitude of the data point in the sub-window after the first move is calculated. This variance value is used as the variance value of the second data point in the moving variance sequence. The above scheme is executed for each move until the sub-window moves out of the channel state information window sequence, thus obtaining the completed moving variance sequence.

[0098] In some embodiments of the present invention, the step of dividing the channel state information sequence into windows according to time order using a sliding window method to obtain a channel state information window sequence further includes:

[0099] The data points collected in this sliding window event are obtained using a sliding window method.

[0100] Obtain the data points after the time point of the back-end segmentation point in the previous channel state information window sequence;

[0101] The data points collected in this sliding window are merged with the data points after the time point of the back-end segmentation point in the previous channel state information window sequence to obtain the current channel state information window sequence.

[0102] In some embodiments of the present invention, if the length of the sequence acquired in this sliding window is L new ， The length of the sequence following the time point of the last segmentation point in the previous channel state information window sequence is L. rest then The length of the channel state information window sequence is L. sw L sw =L rest +L new The length of the sequence is the number of data points in the image, and each time point in the sequence corresponds to one data point. Each time point can be every ms.

[0103] In some embodiments of the present invention, L sw No more than 1.5L new .

[0104] like Figure 2 , 3 As shown, in some embodiments of the present invention, the method further includes the following steps:

[0105] Step S500: Obtain the segmented sequence obtained by segmenting the channel state information sequence based on the front-end segmentation point and the back-end segmentation point; input the parameters in the segmented sequence into a preset classifier; and determine the action type based on the output parameters of the classifier.

[0106] In some embodiments of the present invention, the parameters in the segmentation sequence can be the magnitude of data points, the classifier can be a random forest classifier, and the preset classifier is a pre-trained classifier.

[0107] In the specific implementation process, the output parameter of the classifier can be 1, 2 or 3, etc. If the output parameter of the classifier is 1, it matches the action type corresponding to 1; if the output parameter of the classifier is 2, it matches the action type corresponding to 2; if the output parameter of the classifier is 3, it matches the action type corresponding to 3.

[0108] The type of action can be waving, raising a foot, or kicking, etc.

[0109] Existing threshold-based activity segmentation methods require users to set thresholds or threshold-related parameters based on the granularity of the activity (coarse-grained or fine-grained), which relies on user experience. When the type of activity changes, the original user settings may no longer be applicable, leading to a significant degrade in segmentation performance and ultimately affecting activity recognition performance. Each setting can only be applied to activities of the same granularity, which greatly limits the applicability of existing threshold-based activity extraction methods.

[0110] While existing deep learning-based activity segmentation methods can perceive activity granularity and thus be applicable to activities of varying granularities, they incur significant overhead and more cumbersome processing steps compared to threshold-based methods. Deep learning-based segmentation methods require substantial upfront data collection, and determining the start and end points of the target activity necessitates additional manual labeling. Furthermore, the execution of the deep learning model and preprocessing require substantial computational resources, resulting in excessive time and memory consumption. Therefore, deep learning-based methods are difficult to apply to online activity recognition systems.

[0111] To achieve low overhead, this invention employs a fast preprocessing strategy to remove redundant information from CSI data and designs a core algorithm for activity extraction with low time complexity. The core algorithm includes granularity-aware thresholding and abnormal peak removal.

[0112] For the amplitude of CSI data, the changes in the active portion are much larger than those in the static portion. We found that coarse-grained activity leads to greater changes than fine-grained activity, and the degree of change can be described by variance. We provide a one-dimensional preprocessed CSI that reflects the activity granularity using moving variance in the core algorithm. In this algorithm, a granularity-aware threshold is determined online based on the activity granularity. During the linear traversal of the target activity, we consider that if the activity does not change within a certain distance, and the threshold is fixed and not updated before processing the next activity, then the maximum value of the activity has been found. We use this threshold to find the endpoint and start point sequentially. To avoid false detections of activity start and endpoint caused by interference, we propose anomaly peak removal to distinguish between activity and interference using the duration of the activity. After preprocessing, the activity characteristics will be clearly displayed regardless of the environment. The combination of preprocessing and the core algorithm effectively solves the environment dependency problem.

[0113] This solution implements a threshold-based, low-overhead method for activity signal segmentation at multiple granularities. To meet the low-overhead requirements for online operation, we refocus on thresholding algorithms and design a core algorithm incorporating granularity-aware thresholding for activity signal segmentation. This solution fully meets the low-latency requirements for online operation and is granularity-aware, eliminating the need for users to adjust empirical parameters based on application scenarios, thus enabling activity signal segmentation at different granularities.

[0114] The beneficial effects of this plan include:

[0115] 1. This solution achieves low overhead and fast granularity perception. It adopts a fast preprocessing strategy to remove redundant information and a low-complexity activity extraction core algorithm to achieve low overhead. The core algorithm can automatically perceive the activity granularity and calculate a suitable threshold during the linear traversal of the one-dimensional preprocessed CSI, thus achieving fast granularity perception.

[0116] 2. Environment Independence: After preprocessing, CSI will clearly show the activity characteristics regardless of the environment. The combination of preprocessing, dimensionality reduction and core algorithm effectively solves the environment dependency problem.

[0117] 3. To avoid false detection of activity start and end points caused by interference, we propose abnormal peak removal to use the duration of the activity to distinguish between activity and interference.

[0118] For each step of this invention, we have adopted a specialized design to reduce overhead and innovatively proposed the concept of granularity-aware thresholds. This ensures accurate segmentation of activity signals at different granularities while minimizing processing time, providing users with the best experience. Compared with existing thresholding schemes and deep learning schemes, this invention achieves significant performance improvements in both the accuracy of multi-granularity activity signal segmentation and processing overhead (processing time, memory usage), contributing to the practical application of Wi-Fi activity recognition technology.

[0119] This invention also provides a low-latency activity signal segmentation device for online Wi-Fi sensing. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0120] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned low-latency activity signal segmentation method for online Wi-Fi awareness. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0121] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0122] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0123] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An online wifi-aware low-latency activity signal segmentation method, characterized in that, The steps of the method comprise: Collecting channel state information of a wifi signal by using a network card, constructing a plurality of data points in the channel state information into a channel state information sequence, and dividing the channel state information sequence into channel state information window sequences in time sequence by using a sliding window; Performing dimension reduction on the channel state information window sequences based on a principal component analysis method, and processing the channel state information window sequences after the dimension reduction based on a moving variance algorithm to obtain a moving variance sequence; Extracting each data point in the moving variance sequence in time parameter sequence, taking a variance value of the data point as a calibration variance value if variance values of a first preset number of data points after the data point are all less than a product of the variance value of the data point and a preset ratio parameter, and there is no data point before the data point with a variance value greater than the variance value of the data point; Calculating a granularity perception threshold based on the calibration variance value, calculating the granularity perception threshold according to a preset ratio parameter by using the following formula: granularity perception threshold = calibration variance value * ratio parameter, and calculating a front-end segmentation point and a rear-end segmentation point in the moving variance sequence based on the granularity perception threshold, matching a data point with a first variance value less than the granularity perception threshold in time parameter sequence from the data point corresponding to the calibration variance value, and taking the matched data point as the rear-end segmentation point, matching a data point with a first variance value greater than the granularity perception threshold in time parameter sequence from a starting data point, and taking the matched data point as the front-end segmentation point, and segmenting the channel state information sequence based on the front-end segmentation point and the rear-end segmentation point.

2. The online wifi-aware low-latency activity signal segmentation method of claim 1, wherein, In the step of matching a data point with a first variance value greater than the granularity perception threshold in time parameter sequence from a starting data point, collecting a second preset number of data points after the data point with the first variance value greater than the granularity perception threshold, taking the data point with the first variance value greater than the granularity perception threshold as the front-end segmentation point if a proportion of data points with variance values greater than the granularity perception threshold in the second preset number of data points is greater than a segmentation threshold, and matching a data point with a next variance value greater than the granularity perception threshold after the data point with the first variance value greater than the granularity perception threshold if the proportion of data points with variance values greater than the granularity perception threshold in the second preset number of data points is not greater than the segmentation threshold, re-executing the above steps until the proportion of data points with variance values greater than the granularity perception threshold in a second preset number of data points after the data point with the next variance value greater than the granularity perception threshold is greater than the segmentation threshold, and taking the data point as the front-end segmentation point.

3. The online wifi-aware low-latency activity signal segmentation method of claim 1, wherein, In the step of matching data points with a first variance value less than the granularity perception threshold value in time sequence according to the time parameter and the data points corresponding to the variance value, the matched data point is taken as a rear-end segmentation point. A third preset number of data points after the data point with the first variance value less than the granularity perception threshold value are collected. If the proportion of data points with a variance value less than the granularity perception threshold value in the third preset number of data points is greater than a segmentation threshold value, the data point with the first variance value less than the granularity perception threshold value is taken as a rear-end segmentation point. If the proportion of data points with a variance value less than the granularity perception threshold value in the third preset number of data points is greater than the segmentation threshold value, a data point with a next variance value less than the granularity perception threshold value is matched after the data point with the first variance value less than the granularity perception threshold value. The above step is re-executed until the proportion of data points with a variance value less than the granularity perception threshold value in a third preset number of data points after the data point with the next variance value less than the granularity perception threshold value is greater than the segmentation threshold value. The data point is taken as a rear-end segmentation point.

4. The online wifi-aware low-latency activity signal segmentation method of claim 1, wherein, The step before the step of reducing dimensionality of the channel state information window sequence based on the principal component analysis method further comprises a preprocessing step of the channel state information window sequence, the preprocessing step comprising an amplitude limiting processing step and a denoising processing step; The amplitude limiting processing step is to delete data points with an amplitude value greater than a preset amplitude threshold value in the channel state information window sequence; The denoising processing step is to denoise the channel state information window sequence by using a discrete wavelet transform.

5. The online wifi-aware low-latency activity signal segmentation method according to claim 4, characterized in that, The preprocessing step comprises a compression processing step, the compression processing step being to compress the channel state information window sequence by using an equidistant sampling method.

6. The online wifi-aware low-latency activity signal segmentation method of claim 1, wherein, The step of processing the reduced dimensionality channel state information window sequence based on the moving variance algorithm to obtain a moving variance sequence comprises, obtaining a preset sub-window size of the moving variance algorithm, sliding the sub-window from a data point at a starting time point of the channel state information window sequence according to time sequence, calculating a variance of amplitude values of data points in the sub-window after each sliding, and taking the calculated variance as a variance value of the earliest time point in the sub-window in the moving variance sequence.

7. The low latency activity signal segmentation method for online wifi-aware according to any one of claims 1-6, characterized in that, The step of dividing the channel state information sequence into windows according to time sequence by using a sliding window further comprises: obtaining data points collected by the current sliding window by using a sliding window method; obtaining data points after a time point of a rear-end segmentation point in a previous channel state information window sequence; merging the data points collected by the current sliding window and the data points after the time point of the rear-end segmentation point in the previous channel state information window sequence to obtain a current channel state information window sequence.

8. An online wifi-aware low-latency activity signal segmentation apparatus, characterized by, The device comprises a computer device comprising a processor and a memory, the memory storing computer instructions, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method according to any one of claims 1-7.

Images