Active learning-based method for training behavior recognition model, terminal device, and storage medium

Abstract

The present invention discloses a method for training a behavior recognition model by active learning, a terminal device, and a storage medium. The method includes: inputting a sample to be trained into the behavior recognition model to determine the behavior recognition result of the sample to be trained, and determining a target training sample according to the behavior recognition result; obtaining a behavior label corresponding to the target training sample; and inputting the target training sample and the behavior label into the behavior recognition model for training to update the behavior recognition model. By inputting the sample to be trained into the behavior recognition model to determine the behavior recognition result of the sample to be trained, and then determining the target training sample according to the behavior recognition result, the present invention selects the target training sample from the samples to be trained as the "difficult samples" of the current behavior recognition model, so as to reduce the participation of invalid or inefficient samples to be trained in the training and improve the training efficiency of the behavior recognition model.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior recognition, and in particular, to a method for training a behavior recognition model with active learning, a terminal device, and a storage medium. Background Art

[0002] In recent years, the technology for training behavior recognition models has developed rapidly and has been gradually applied in various fields such as motion tracking, security monitoring, and medical health. The current methods for training behavior recognition models can be divided into three types: vision-based, wearable sensor-based, and wireless signal-based. Among them, the training of behavior recognition models based on wireless signals such as RFID, ZigBee, and WiFi distinguishes corresponding actions according to the changes in parameters such as signal strength and channel quality of wireless signals caused by different behaviors of users. However, the current training method for behavior recognition models based on signals often uses deep learning in a fully supervised manner. Therefore, a large number of labeled training samples are required to build an identification model to ensure the performance of the model, resulting in high deployment costs and low model training efficiency.

[0003] The above content is only used to assist in understanding the technical solution of the present invention, and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main object of the present invention is to provide a method for training a behavior recognition model with active learning, a terminal device, and a storage medium, aiming to reduce the number of required labeled training samples while taking into account the performance of building an identification model by deep learning, thereby improving the training efficiency of the model and reducing the deployment cost.

[0005] To achieve the above object, a method for training a behavior recognition model with active learning, the method for training a behavior recognition model with active learning includes:

[0006] Input a to-be-trained sample into the behavior recognition model to determine the behavior recognition result of the to-be-trained sample, and determine a target training sample according to the behavior recognition result;

[0007] Obtain a behavior label corresponding to the target training sample;

[0008] Input the target training sample and the behavior label into the behavior recognition model for training to update the behavior recognition model.

[0009] Optionally, the step of determining a target training sample according to the behavior recognition result includes:

[0010] When the behavior recognition result is difficult to determine, determine the to-be-trained sample as the target training sample, or when the probability values corresponding to the limb movement behavior types in the behavior recognition result are all less than or equal to a preset probability value, determine the to-be-trained sample as the target training sample.

[0011] Optionally, the step of obtaining the behavior label corresponding to the target training sample includes:

[0012] Obtain the channel state information data and video information when the target object makes an action;

[0013] Determine the to-be-trained sample according to the channel state information data, where the to-be-trained sample includes the target training sample;

[0014] Determine the correspondence between the to-be-trained sample and the video information according to the time information;

[0015] Obtain the behavior label corresponding to the target training sample according to the correspondence.

[0016] Optionally, the step of determining the to-be-trained sample according to the channel state information data includes:

[0017] Preprocess the channel state information data to obtain the preprocessed channel state information data;

[0018] Generate a channel state information spectrogram according to the preprocessed channel state information data and the Fourier transform algorithm to determine the to-be-trained sample.

[0019] Optionally, the step of preprocessing the channel state information data to obtain the preprocessed channel state information data includes:

[0020] Perform a first denoising process on the channel state information data through a preset filter to obtain the denoised channel state information data;

[0021] Perform a second denoising process on the denoised channel state information data according to the principal component analysis method to obtain the preprocessed channel state information data.

[0022] Optionally, after the step of inputting the target training sample and the behavior label into the behavior recognition model for training to update the behavior recognition model, it further includes:

[0023] Obtain the channel state information data when the target object makes an action;

[0024] Determine the limb behavior action of the target object according to the channel state information data and the updated behavior recognition model.

[0025] Optionally, the to-be-trained sample includes data samples corresponding to the same limb movement behaviors in multiple application environments.

[0026] Optionally, the network model corresponding to the behavior recognition model adopts the AlexNet model.

[0027] In addition, to achieve the above object, the present invention further provides a terminal device, which includes: a memory, a processor, and an active learning behavior recognition model training program stored in the memory and executable on the processor. When the active learning behavior recognition model training program is executed by the processor, it implements each step of the active learning behavior recognition model training method as described above.

[0028] In addition, to achieve the above object, the present invention further provides a storage medium, on which an active learning behavior recognition model training program is stored. When the active learning behavior recognition model training program is executed by the processor, it implements each step of the active learning behavior recognition model training method as described above.

[0029] The active learning behavior recognition model training method, terminal device, and storage medium proposed by the present invention determine the behavior recognition result of the to-be-trained sample by inputting the to-be-trained sample into the behavior recognition model, and then determine the target training sample according to the behavior recognition result, so as to select the target training sample from the to-be-trained samples as the "difficult sample" of the current behavior recognition model. While reducing the participation of ineffective or inefficient to-be-trained samples in training and improving the training efficiency of the behavior recognition model, the training sample set of the behavior recognition model can be enriched, making the training sample set of the behavior recognition model diverse. Then, the behavior recognition model is trained with the target training sample and the behavior label corresponding to the target training sample to update the behavior recognition model, improving the recognition performance of the updated behavior recognition model, enhancing the recognition accuracy, and enhancing the robustness, universality, and practicality of the behavior recognition model. Description of the Drawings

[0030] Figure 1 It is a schematic structural diagram of a terminal device involved in each embodiment of the active learning behavior recognition model training method of the present invention;

[0031] Figure 2 It is a schematic flowchart of the first embodiment of the active learning behavior recognition model training method of the present invention;

[0032] Figure 3 It is a simple schematic diagram of the wireless signal sensing area;

[0033] Figure 4 It is a schematic flowchart of determining the target training sample in the first embodiment of the active learning behavior recognition model training method of the present invention;

[0034] Figure 5This is a performance comparison chart of the training of the present invention using target training samples and the training of behavior recognition based on deep learning using all samples;

[0035] Figure 6 This is the network structure diagram of the AlexNet model;

[0036] Figure 7 This is a schematic flowchart of the second embodiment of the training method of the active learning behavior recognition model of the present invention;

[0037] Figure 8 This is a schematic diagram of video information corresponding to two kinds of limb behaviors;

[0038] Figure 9 This is a simple schematic flowchart of the training method of the active learning behavior recognition model of the present invention.

[0039] The realization, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific Embodiments

[0040] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] The present invention provides a training method for an active learning behavior recognition model. The training method for the active learning behavior recognition model includes:

[0042] Input the training samples to be trained into the behavior recognition model to determine the behavior recognition result of the training samples to be trained, and determine the target training samples according to the behavior recognition result;

[0043] Obtain the behavior labels corresponding to the target training samples;

[0044] Input the target training samples and the behavior labels into the behavior recognition model for training to update the behavior recognition model.

[0045] The training method for the active learning behavior recognition model of the present invention determines the behavior recognition result of the training sample to be input into the behavior recognition model, and then determines the target training sample according to the behavior recognition result, so as to select the target training sample from the training samples to be used as the "difficult sample" of the current behavior recognition model. While reducing the participation of ineffective or inefficient training samples in training and improving the training efficiency of the behavior recognition model, it also reduces the deployment cost, enriches the training sample set of the behavior recognition model, makes the training sample set of the behavior recognition model diverse, and then trains the behavior recognition model through the target training sample and the behavior label corresponding to the target training sample to update the behavior recognition model, improving the recognition performance of the updated behavior recognition model and the recognition accuracy, and enhancing the robustness, universality and practicability of the behavior recognition model.

[0046] In the following description, the suffixes such as "module", "component" or "unit" used to represent components are only for the convenience of the description of the present invention, and they have no specific meaning by themselves. Therefore, "module", "component" or "unit" can be used interchangeably.

[0047] Please refer to Figure 1 , Figure 1 FIG. is a schematic structural diagram of a terminal device involved in various embodiments of the training method for the active learning behavior recognition model of the present invention. Among them, the terminal device involved in the training method for the active learning behavior recognition model of the present invention is implemented in various forms. For example, the terminal device described in the present invention may include terminal devices such as servers, mobile phones, tablet computers, laptop computers, handheld computers, and personal digital assistants (Personal Digital Assistant, PDA).

[0048] As Figure 1 shown, the terminal device may include: a memory 101 and a processor 102. Those skilled in the art can understand that Figure 1 the block diagram of the terminal shown does not constitute a limitation on the terminal. The terminal may include more or fewer components than shown, or combine some components, or have different component arrangements. Among them, the memory 101 stores an operating system and a training program for the active learning behavior recognition model. The processor 102 is the control center of the terminal device. The processor 102 executes the training program for the active learning behavior recognition model stored in the memory 101 to implement the steps of various embodiments of the training method for the active learning behavior recognition model of the present invention.

[0049] Optionally, the terminal device may further include a communication unit 103. The communication unit 103 establishes data communication with other terminal devices through a network protocol (this data communication may be IP communication, WiFi communication, or a Bluetooth channel). Exemplarily, other terminal devices such as an auxiliary vision device, where the auxiliary vision device is used to obtain video information of a target object such as a person's actions, and the video information can be sent through the auxiliary vision device to obtain the video information of the target object's actions.

[0050] Optionally, the terminal device is a wireless signal receiving terminal device.

[0051] Based on the above structural block diagram of the terminal device, various embodiments of the active learning behavior recognition model training method of the present invention are proposed.

[0052] In the first embodiment, the present invention provides an active learning behavior recognition model training method. Please refer to Figure 2 , Figure 2 which is a schematic flowchart of the first embodiment of the active learning behavior recognition model training method of the present invention. In this embodiment, the active learning behavior recognition model training method includes the following steps:

[0053] Step S10, determining a target training sample according to the sample to be trained;

[0054] Step S20, obtaining a behavior label corresponding to the target training sample;

[0055] Step S30, inputting the target training sample and the behavior label into the behavior recognition model for training to update the behavior recognition model.

[0056] The sample to be trained refers to the data sample corresponding to the channel state information data collected when the target object is acting in the wireless signal sensing area. The behavior label refers to the limb behavior actions of the target object when acting in the wireless signal sensing area. For example, the behavior label includes, but is not limited to, the identifications of limb behavior actions such as walking, jumping, turning around, squatting, standing up, and running, and the limb behavior actions of the target object can be determined through the behavior label.

[0057] Optionally, before step S10, it includes: obtaining the channel state information data when the target object is acting; determining the sample to be trained according to the channel state information data.

[0058] The target object refers to the object to be monitored. Optionally, the target object can be a person. The channel state information (CSI) data refers to the channel state information data of the wireless signal physical layer, mainly including the frequency domain information of the signal, such as amplitude and phase, etc.

[0059] In the actual application process, with the development of wireless networks, wireless networks have gradually achieved full coverage. When wireless signals pass through obstacles such as moving people, reflections, scattering, etc. occur, forming multipath superposition signals. By analyzing the corresponding changes in wireless signals caused by human behavior, different movements of the human body have different impacts on the wireless signal channel state information. The terminal device can collect the information on the impact of the movement of the target object within the wireless signal sensing area on the wireless channel state by using a high-performance commercial wireless network card, so as to obtain the channel state information data when the target object moves.

[0060] Exemplarily, to collect the information on the impact of human movement within the wireless signal sensing area on the wireless channel state to obtain the channel state information data when the human moves, it can be achieved by using a high-performance commercial wireless network card to collect the information on the impact of human movement within the wireless signal sensing area on the wireless channel state to obtain the channel state information data when the human moves.

[0061] For the specific implementation of determining the training samples according to the channel state information data, please refer to the second embodiment section, which will not be elaborated in detail in this embodiment.

[0062] Optionally, the wireless signal channel state can be the Wi-Fi channel state.

[0063] It should be noted that the wireless signal sensing area can be determined by the sensing area formed by the wireless signal transmitting terminal device and the wireless signal receiving terminal device.

[0064] Optionally, the number of transmitting antennas configured by the wireless signal transmitting terminal device can be one or at least two.

[0065] Optionally, the number of receiving antennas configured by the wireless signal receiving terminal device can be one or at least two.

[0066] Optionally, the wireless signal transmitting terminal device uses a WIFI transmitting terminal device, and the wireless signal receiving terminal device uses a WIFI receiving terminal device.

[0067] Exemplarily, a sensing area can be formed by setting up a WiFi transmitting terminal device (configured with 1 omnidirectional antenna) and a WiFi receiving terminal device (configured with 3 omnidirectional antennas) to determine the wireless signal sensing area. For details, please refer to Figure 3 , Figure 3 which is a simple schematic diagram of the wireless signal sensing area.

[0068] As an alternative implementation, please refer to Figure 4 , Figure 4Schematic diagram of the process for determining the target training samples in the first embodiment of the method for training an active learning behavior recognition model according to the present invention. Step S10 includes:

[0069] Step S11: Input the sample to be trained into the behavior recognition model to determine the behavior recognition result of the sample to be trained.

[0070] Step S12: Determine the target training samples according to the behavior recognition result.

[0071] The behavior recognition result includes at least one of difficult to distinguish, recognizable, limb movement behavior type, and the probability value corresponding to the limb movement behavior type. Among them, difficult to distinguish can be understood as unable to recognize. The limb movement behavior type refers to limb behavior actions, which include but are not limited to walking, jumping, turning, squatting, standing up, and running. The probability value corresponding to the limb movement behavior type refers to the probability value of the limb movement behavior obtained by recognizing the sample to be trained. Exemplarily, input the sample to be trained into the behavior recognition model, and the behavior recognition results of the sample to be trained for walking, jumping, turning, squatting, standing up, and running are respectively: 0.1, 0.1, 0.1, 0, 0.5, and 0.2, indicating that the behavior label or limb movement behavior type or limb movement behavior corresponding to the sample to be trained is standing up.

[0072] Optionally, step S12 includes:

[0073] When the behavior recognition result is difficult to distinguish, determine the sample to be trained as the target training sample; or when the probability values corresponding to the limb movement behavior types in the behavior recognition result are all less than or equal to a preset probability value, determine the sample to be trained as the target training sample.

[0074] When the behavior recognition result is difficult to distinguish, it indicates that the sample to be trained is a "difficult sample" for the behavior recognition model. Similarly, when the probability values corresponding to the limb movement behavior types in the behavior recognition result are all less than or equal to the preset probability value, it indicates that the current behavior recognition model has low accuracy in recognizing the sample to be trained, which also indicates that the sample to be trained is a "difficult sample" for the behavior recognition model. Furthermore, determine the sample to be trained as the target training sample to train the behavior recognition model according to the target training sample, so that the behavior recognition model trained according to the target training sample can recognize the target training sample or improve the accuracy of recognizing the target training sample.

[0075] It is easy to understand that when the probability value corresponding to the limb movement behavior type in the behavior recognition result is greater than the preset probability value, it indicates that the limb movement behavior corresponding to the to-be-trained sample can be accurately recognized by the behavior recognition model, and the to-be-trained sample can be reduced or not used to train the behavior recognition model.

[0076] It should be noted that in this embodiment, the method of using active learning to select "difficult samples" is adopted to preferentially select samples that are difficult for the current behavior recognition model to distinguish or have low accuracy in behavior recognition from the to-be-trained samples as target training samples. By using the target samples rich in information to train the behavior recognition model, the participation of invalid or inefficient samples in training is reduced, so as to reduce the demand for to-be-trained samples with behavior labels, and further achieve the actual application that meets the cross-environment requirements.

[0077] Optionally, the to-be-trained samples include data samples corresponding to the same limb movement behaviors in multiple application environments. The limb movement behaviors include, but are not limited to, walking, jumping, turning, squatting, standing up, and running. Exemplarily, it is assumed that the application environments include 6 daily environments such as corridors, meeting rooms, laboratories, halls, elevator entrances and exits, and open platforms. Among them, a wireless signal sensing area is set in each application environment. For the acquisition of to-be-trained samples, 4 limb behavior actions such as walking, jumping, turning, and squatting and standing up can be carried out in each daily environment by recruiting volunteers, and the channel state information data corresponding to the 4 limb behavior actions in each daily environment is collected. Then, the to-be-trained samples are determined according to the channel information data to enrich the training sample set of the behavior recognition model, so that the training sample set of the behavior recognition model has diversity, and further achieve the actual application that meets the cross-environment requirements.

[0078] Exemplarily, please refer to Figure 5 , Figure 5 is a performance comparison chart of training the behavior recognition based on deep learning using all samples and training with the target training samples in the present invention. Among them, training with the target training samples includes training the behavior recognition model with the data sets of 6 environments. In the present invention, training with 15% of the labeled samples can achieve an identification accuracy of 58.97%. Compared with the behavior recognition system based on deep learning, which can only achieve an identification accuracy of 62.19% when training with 100% of the to-be-trained samples with behavior labels, when achieving an identification accuracy similar to that of the behavior recognition system based on deep learning, the active learning adopted in this embodiment selects target training samples from the to-be-trained samples for training, and the number of samples with behavior labels required can be reduced by more than 80%.

[0079] Exemplarily, samples that are difficult for the current behavior recognition model to distinguish or have a relatively low accuracy in behavior recognition are preferentially selected from the samples to be trained as target training samples, which can be regarded as "difficult samples". Taking the animal recognition model to distinguish cats and dogs as an example: when the animal recognition model can already accurately distinguish common types of cats and dogs, such as Huskies, Golden Retrievers, Ragdoll cats, and Tabby cats, there is no need to add such data samples to the animal recognition model. For the animal recognition model, it is difficult to distinguish or accurately identify robotic cats and cartoon dogs, so the data samples corresponding to robotic cats and cartoon dogs respectively are "difficult samples" relative to the animal recognition model.

[0080] Optionally, in step S20 of obtaining the behavior label corresponding to the target training sample, the channel state information data and video information when the target object makes an action can be obtained. The sample to be trained is determined according to the channel state information data, the correspondence between the sample to be trained and the video information is determined according to the time information, and the behavior label corresponding to the target training sample is obtained according to the correspondence. Among them, the sample to be trained includes the target training sample. For the specific implementation, reference can be made to the second embodiment.

[0081] Among them, obtaining the channel state information data and video information when the target object makes an action, and determining the sample to be trained according to the channel state information data can be used in the actual application scenario or actual application environment. When the behavior recognition model has difficulty distinguishing the data sample corresponding to the channel state information data obtained when the target object makes an action, or when the accuracy of the data sample corresponding to the channel state information data obtained when the behavior recognition model recognizes the target object's action is relatively low, the data sample corresponding to the channel state information data can be determined as the target training sample to enrich the training sample set of the behavior recognition model, so that the training sample set of the behavior recognition model has diversity. The behavior recognition model is trained with the target training sample to update the behavior recognition model and improve the recognition performance and accuracy of the updated behavior recognition model. It should be noted that the sample to be trained includes the target training sample.

[0082] Optionally, in step S20 of obtaining the behavior label corresponding to the target training sample, the target data sample matching the target training sample can be found from the preset data sample set, and then the behavior label corresponding to the target data sample is determined based on the preset data sample set to obtain the behavior label corresponding to the target training sample. Among them, the preset data sample set includes a large number of data samples and the corresponding relationship between the data samples and the behavior labels. Among them, the data samples can be directly obtained by searching on a search engine.

[0083] Step S30: Input the target training samples and behavior labels into the behavior recognition model for training to update the behavior recognition model, so as to improve the recognition performance of the updated behavior recognition model and the recognition accuracy.

[0084] Optionally, the number of target training samples is multiple. When inputting the target training samples and behavior labels into the behavior recognition model for training to update the behavior recognition model, the training of the behavior recognition model can be stopped when the probability value of the limb movement behavior type in the behavior recognition result obtained by the behavior recognition model for the target training samples is greater than the preset probability value; or the training of the behavior recognition model can also be completed when all multiple target training samples have been used for training the behavior recognition model.

[0085] As an optional implementation manner, after step S30, it further includes:

[0086] Obtain the channel state information data when the target object makes an action;

[0087] Determine the limb behavior actions of the target object according to the channel state information data and the updated behavior recognition model.

[0088] To obtain the channel state information data when the target object makes an action, the terminal device can collect the information on the impact of the movement of the target object within the wireless signal sensing area on the wireless channel state through a high-performance commercial wireless network card, so as to obtain the channel state information data when the target object makes an action.

[0089] To determine the limb behavior actions of the target object according to the channel state information data and the updated behavior recognition model, the channel state information spectrogram can be determined according to the channel state information data, and the limb behavior actions of the target object can be determined according to the channel state information spectrogram and the updated behavior recognition model.

[0090] Among them, the step of determining the channel state information spectrogram according to the channel state information data includes preprocessing the channel state information data to obtain the preprocessed channel state information data, and generating the channel state information spectrogram according to the preprocessed channel state information data and the Fourier transform algorithm. The specific implementation of this step can refer to the second embodiment and will not be specifically described in this embodiment.

[0091] Optionally, the network model corresponding to the behavior recognition model adopts the AlexNet model. Please refer to Figure 6 , Figure 6 which is the network structure diagram of the AlexNet model.

[0092] In the technical solution disclosed in this embodiment, the to-be-trained samples are input into the behavior recognition model to determine the behavior recognition results of the to-be-trained samples, and then the target training samples are determined according to the behavior recognition results, so as to select the target training samples from the to-be-trained samples as the "difficult samples" of the current behavior recognition model, reducing the participation of ineffective or inefficient to-be-trained samples in training, improving the training efficiency of the behavior recognition model, reducing the deployment cost, enriching the training sample set of the behavior recognition model, making the training sample set of the behavior recognition model diverse, and then training the behavior recognition model through the target training samples and the behavior labels corresponding to the target training samples to update the behavior recognition model, improving the recognition performance of the updated behavior recognition model and the recognition accuracy, and enhancing the robustness, universality and practicability of the behavior recognition model.

[0093] In the second embodiment proposed on the basis of the first embodiment, please refer to Figure 7 , Figure 7 which is a schematic flowchart of the second embodiment of the active learning-based behavior recognition model training method of the present invention. In this embodiment, step S20 includes:

[0094] Step S21, obtaining the channel state information data and video information when the target object makes an action;

[0095] Step S22, determining the to-be-trained samples according to the channel state information data, where the to-be-trained samples include the target training samples;

[0096] Step S23, determining the correspondence between the to-be-trained samples and the video information according to the time information;

[0097] Step S24, obtaining the behavior labels corresponding to the target training samples according to the correspondence.

[0098] Based on the first embodiment, information on the influence of the movement of the target object in the wireless signal sensing area on the wireless channel state can be collected by using a high-performance commercial wireless network card to obtain the channel state information data when the target object makes an action.

[0099] Optionally, auxiliary vision devices can be set in the wireless signal sensing area to obtain video information when the target object, such as a person, makes an action. Exemplarily, the auxiliary vision device is a camera.

[0100] Optionally, the way to obtain the channel state information data and video information when the target object makes an action is to obtain them in real time.

[0101] Optionally, the channel state information data and video information when the target object performs an action are the channel state information data and video information synchronized when the target object performs the action obtained based on the arrangement order of time information, that is, there is a synchronous correspondence relationship between the channel state information data and the video information at the same time point or the same time period.

[0102] In the actual application process, when the wireless signal receiving terminal device receives a wireless signal, it will receive multiple data packets. Each data packet can include time information and channel state information data. Among them, the channel state information data in the data packet can be arranged according to the time information to form a CSI time series. Similarly, the auxiliary vision device can collect the video information when the target object performs an action. Among them, the video information contains time information. In this way, the synchronized channel state information data and video information can be obtained corresponding to the same time information such as time point or time period, so as to determine the correspondence relationship between the channel state information data and the video information. On the basis of determining the training sample to be determined according to the channel state information data, the correspondence relationship between the training sample to be determined and the video information is further determined.

[0103] As an optional implementation manner, step S50 includes:

[0104] Perform preprocessing on the channel state information data to obtain the preprocessed channel state information data;

[0105] Generate a channel state information spectrogram according to the preprocessed channel state information data and the Fourier transform algorithm to determine the training sample to be determined.

[0106] Optionally, performing preprocessing on the channel state information data to obtain the preprocessed channel state information data includes:

[0107] Perform first denoising processing on the channel state information data through a preset filter to obtain the denoised channel state information data;

[0108] Perform second denoising processing on the denoised channel state information data according to the principal component analysis method to obtain the preprocessed channel state information data.

[0109] In actual applications, due to problems with the wireless signal receiving terminal device or the external environment, there are outliers in the channel state information data in the CSI time series of the acquired wireless signal, which affects the accuracy of extracting the behavioral characteristics of the target object. To accurately obtain the behavioral characteristics of the target object, the collected channel state information data can be subjected to first denoising processing through a preset filter to obtain the denoised channel state information data.

[0110] Optionally, the preset filter may adopt a Hampel filter to perform first denoising processing on the channel state information data through the Hampel filter to obtain the denoised channel state information data. Exemplarily, the Hampel filter may be used to detect abnormal data in the channel state information data. By determining the median value in the collected channel state information data, the difference between each data in the channel state information data and the median value is determined. When the absolute value of the difference is greater than or equal to a preset value, the data corresponding to the difference is determined as abnormal data, and the abnormal data is replaced with the median value to avoid loss or missing of the channel state information data. Alternatively, the Hampel filter can also identify the position where the CSI time series, i.e., the abnormal value of the channel state information data, appears, and adopt a least squares support vector machine regression model to detect the abnormal value in the CSI time series based on the recursive prediction method to realize the analysis and processing of monitoring the abnormal information value in the CSI time series. Similarly, in practical applications, after deleting the abnormal value at a certain position in the CSI time series, it may cause partial loss or missing of the CSI time series, disrupting the corresponding relationship between the channel state information data and the time information, and thus affecting the entire CSI time series. Interpolation processing can be performed at the position where the information abnormal value is deleted in the CSI time series to supplement the missing data and ensure the accuracy of the information sequence of the CSI time series.

[0111] Perform second denoising processing on the denoised channel state information data according to the principal component analysis method to obtain the preprocessed channel state information data. The principal component analysis method can be used to further process the denoised channel state information data. The denoised channel state information data can be regarded as the first CSI data. Each first CSI data is divided into multiple sub-CSI data with the same time length, such as multiple sub-CSI data with a length of 8 seconds each. Principal component analysis is performed on each sub-CSI data respectively, and then the top 20 principal components in the ranking of each first CSI data are selected for normalization to eliminate data noise, extract the main features of the data, and reduce the dimension of the data space.

[0112] Optionally, the specific implementation of performing second denoising processing on the denoised channel state information data according to the principal component analysis method may also refer to other implementable ways of the existing technical solutions, and this embodiment does not make specific limitations on this step.

[0113] Generate a channel state information spectrogram according to the preprocessed channel state information data and the Fourier transform algorithm to determine the training samples to be used. Among them, through a preset Fourier transform algorithm such as the short-time Fourier transform algorithm, the preprocessed channel state information data is transformed into a CSI spectrogram. The preset Fourier transform algorithm is:

[0114]

[0115] Wherein, x[n] is the discrete-time sequence of the signal, and w[m] is the window of the short-time Fourier transform, and the spectrogram generated by the square of its amplitude is:

[0116] Spectrogram{x[n]} = |X(e jw , n)| 2

[0117] As an alternative implementation, step S70 determines the behavior label corresponding to the target training sample according to the corresponding relationship, based on the corresponding relationship between the training samples to be determined sorted based on time information and the video information, wherein the training samples to be determined include the target training sample, and the time information where the training sample to be determined identical to the target training sample is located can be found through the determined target training sample, and then the video information corresponding to the time information is obtained, and the behavior label corresponding to the target training sample is determined according to the video information.

[0118] Optionally, to determine the behavior label corresponding to the target training sample according to the video information, the video information is manually identified to determine the behavior label corresponding to the target training sample through the video information, and then the corresponding relationship between the target training sample and the behavior label is established. Exemplarily, the corresponding relationship between the target training sample and the behavior label can be manually input through the training interface to obtain the behavior label corresponding to the target training sample. Exemplarily, reference can be made to Figure 8 , Figure 8 which are schematic diagrams of the video information corresponding to two limb behaviors respectively.

[0119] Exemplarily, reference can be made to Figure 9 , Figure 9 which is a simple flowchart of the method for training the behavior recognition model of active learning according to the present invention.

[0120] Optionally, to determine the behavior label corresponding to the target training sample according to the video information, a video information recognition algorithm can be used to actively analyze the video information to obtain the behavior label corresponding to the video information, so as to determine the behavior label corresponding to the target training sample.

[0121] Optionally, the time information can be a time point or a time period.

[0122] The present invention also provides a terminal device, which includes: a memory, a processor, and an active learning behavior recognition model training program stored in the memory and executable on the processor. When the active learning behavior recognition model training program is executed by the processor, the steps of the method for training the active learning behavior recognition model in any of the above embodiments are implemented.

[0123] The present invention also provides a storage medium, on which a training program for an active learning behavior recognition model is stored. When the training program for the active learning behavior recognition model is executed by a processor, the steps of the training method for the active learning behavior recognition model described in any of the above embodiments are implemented.

[0124] In the embodiments of the terminal device and the storage medium provided by the present invention, all the technical features of the above embodiments of the training method for the active learning behavior recognition model are included. The content of the specification expansion and explanation is basically the same as that of the above embodiments of the training method for the active learning behavior recognition model, and will not be repeated here.

[0125] It should be noted that in this article, the term "comprising", "including" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or system comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or system. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or system comprising such element.

[0126] The serial numbers of the above embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments.

[0127] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, a controlled terminal, or a network device, etc.) to execute the methods of each embodiment of the present invention.

[0128] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A training method for an active learning behavior recognition model, characterized in that, The training method of the behavior recognition model based on active learning includes: Inputting the training samples to be trained into the behavior recognition model to determine the behavior recognition result of the training samples to be trained, and determining the target training samples according to the behavior recognition result; Obtaining the channel state information data and video information when the target object makes an action; Determining the training samples to be trained according to the channel state information data, where the training samples to be trained include the target training samples; Determining the corresponding relationship between the training samples to be trained and the video information according to the time information; Obtaining the behavior labels corresponding to the target training samples according to the corresponding relationship; Inputting the target training samples and behavior labels into the behavior recognition model for training to update the behavior recognition model; The step of determining the training samples to be trained according to the channel state information data includes: Preprocessing the channel state information data to obtain the preprocessed channel state information data; Generating a channel state information spectrogram according to the preprocessed channel state information data and the Fourier transform algorithm to determine the training samples to be trained.

2. The method for training an active learning-based behavior recognition model according to claim 1, wherein, The step of determining the target training samples according to the behavior recognition result includes: When the behavior recognition result is difficult to distinguish, determining the training samples to be trained as the target training samples, or when the probability values corresponding to the limb movement behavior types in the behavior recognition result are all less than or equal to the preset probability value, determining the training samples to be trained as the target training samples.

3. The method for training an active learning behavior recognition model according to claim 1, characterized in that The step of preprocessing the channel state information data to obtain the preprocessed channel state information data includes: Performing first denoising processing on the channel state information data through a preset filter to obtain the denoised channel state information data; Performing second denoising processing on the denoised channel state information data according to the principal component analysis method to obtain the preprocessed channel state information data.

4. The method for training an active learning-based behavior recognition model according to claim 1, wherein After the step of inputting the target training samples and behavior labels into the behavior recognition model for training to update the behavior recognition model, it further includes: Obtaining the channel state information data when the target object makes an action; Determining the limb behavior actions of the target object according to the channel state information data and the updated behavior recognition model.

5. The method for training an active learning-based behavior recognition model according to claim 1, wherein The training samples to be trained include data samples corresponding to the same limb movement behaviors in multiple application environments.

6. The method for training an active learning-based behavior recognition model according to claim 1, wherein The network model corresponding to the behavior recognition model adopts the AlexNet model.

7. An end device for training a behavior recognition model with active learning, characterized in that, The terminal device for training the behavior recognition model based on active learning includes: a memory, a processor, and an active learning behavior recognition model training program stored in the memory and executable on the processor. When the active learning behavior recognition model training program is executed by the processor, it realizes the steps of the active learning behavior recognition model training method according to any one of claims 1-6.

8. A storage medium, characterized in that, An active learning behavior recognition model training program is stored on the storage medium. When the active learning behavior recognition model training program is executed by the processor, it realizes the steps of the active learning behavior recognition model training method according to any one of claims 1-6.

Images