Gesture recognition method, device, storage medium and wearable device

Abstract

This application relates to the field of biosignal processing technology, and discloses a gesture recognition method, device, storage medium, and wearable device. The method includes: collecting a user's gesture to be calibrated to obtain calibration gesture data; performing a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector; calibrating a preset action category based on the calibration action hypervector; and recognizing the user's gesture based on the calibrated preset action category. Compared to existing methods that upload low-dimensional gesture data to a cloud server for calibration, this application can first perform a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector, and then directly perform calibration locally based on the calibration action hypervector, without uploading to a server, thereby reducing calibration time.

Description

Technical Field

[0001] This application relates to the field of biosignal processing technology, and in particular to a gesture recognition method, apparatus, storage medium and wearable device. Background Technology

[0002] Gesture recognition is an important means of achieving human-computer interaction. Currently, many methods such as electromyography (EMG), inertial measurement unit (IMU), photoplethysmography (PPG), and ultrasound can achieve gesture recognition. Most existing wearable devices (such as wristband devices) use EMG or IMU. That is, EMG data is collected through EMG electrodes at the wristband or IMU data is collected by IMU sensors. Then, the gesture recognition model deployed inside the device analyzes the data and outputs the gesture recognition result.

[0003] However, when calibrating a gesture recognition model, since the collected EMG or IMU data are low-dimensional gesture data, deep learning algorithms are required to calibrate the model. However, deep learning algorithms involve significant computational demands and long training times, and require high precision and performance from hardware. Current wearable devices often lack the computing power to support direct calibration training on internal chips or field-programmable gate arrays (FPGAs). Therefore, the common approach is to upload the collected EMG or IMU data to a cloud server, which then calibrates the gesture recognition model using deep learning algorithms and returns the calibrated model parameters to the wearable device, resulting in a lengthy calibration process. Summary of the Invention

[0004] The main objective of this application is to provide a gesture recognition method, apparatus, storage medium, and wearable device, aiming to solve the technical problem that existing gesture recognition model calibration methods use low-dimensional gesture data, which requires uploading the gesture data to a cloud server for calibration, resulting in long calibration times.

[0005] To achieve the above objectives, embodiments of this application provide a gesture recognition method, the method comprising:

[0006] The user's gestures to be calibrated are collected to obtain calibration gesture data;

[0007] The calibration gesture data is subjected to dimensionality upscaling to obtain a calibration action hypervector;

[0008] The preset action category is calibrated based on the calibrated action hypervector, and the user's gesture to be recognized is recognized based on the calibrated preset action category.

[0009] In one embodiment, the step of performing a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector includes:

[0010] Select the entity used in this gesture recognition, the entity including at least one of the sensor group or electrode group and the signal amplitude;

[0011] Determine the entity information corresponding to the entity in the calibration gesture data, wherein the entity information includes entity label information and amplitude signal sequence information;

[0012] The entity information is upgraded using spatial and temporal encoding to obtain the hypervector of the action to be calibrated.

[0013] In one embodiment, the method further includes:

[0014] The entity type of the entity includes a first entity type and a second entity type;

[0015] When the entity is the first entity type, the hypervectors corresponding to the entity information of the entity are quasi-orthogonal, and the entity includes the sensor group and the electrode group;

[0016] When the entity is the second entity type, the hypervectors corresponding to the entity information of the entity have similarity, and the entity includes the electrode group and the signal amplitude.

[0017] In one embodiment, the step of performing dimensionality upscaling on the entity information through spatial and temporal encoding to obtain the action hypervector to be calibrated includes:

[0018] The entity label hypervector corresponding to each entity label information is determined by the first preset mapping relationship;

[0019] Based on the preset binning rules, the signal amplitudes in each amplitude signal sequence information are binned, and the supervector of the amplitude signal sequence corresponding to each binned signal amplitude is determined by the second preset mapping relationship.

[0020] Bind each entity label hypervector to the corresponding amplitude signal sequence hypervector to obtain the bound hypervector of the entity information at each time point;

[0021] The binding hypervectors are superimposed to obtain entity information hypervectors at each time step, and the entity information hypervectors at each time step are time-encoded to obtain calibration action hypervectors.

[0022] In one embodiment, before the step of collecting the user's gesture to be calibrated and obtaining calibration gesture data, the method further includes:

[0023] Determine the entity type of each entity;

[0024] When the entity type is the first entity type, the entity is defined according to a preset hypervector dimension to obtain the entity label hypervector corresponding to the entity;

[0025] Obtain the entity label information corresponding to the entity, and construct a first preset mapping relationship based on the entity label information and the corresponding entity label hypervector.

[0026] In one embodiment, after the step of determining the entity type of each entity, the method further includes:

[0027] When the entity type is the second entity type and the entity is an electrode group, a sub-entity is selected from the entity;

[0028] Define the sub-entity according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the sub-entity;

[0029] The entity label hypervector corresponding to the sub-entity is flipped, and the flipped result is used as the entity label hypervector corresponding to the entity.

[0030] In one embodiment, after the step of determining the entity type of each entity, the method further includes:

[0031] When the entity type is the second entity type and the entity is a signal amplitude, the entity is scaled up and down, and the scaled entity is divided into boxes according to a preset boxing rule.

[0032] The target box is determined based on the results of each sub-box, and the target box is defined according to the preset vector dimension to obtain the amplitude signal sequence supervector corresponding to the target box;

[0033] The amplitude signal sequence supervector corresponding to the target bin is flipped to obtain the amplitude signal sequence supervector corresponding to each binning result;

[0034] Obtain the signal amplitude after binning corresponding to each binning result, and construct a second preset mapping relationship based on the signal amplitude after binning and the corresponding amplitude signal sequence supervector.

[0035] In one embodiment, the step of temporally encoding the entity information hypervectors at each time point to obtain calibration action hypervectors includes:

[0036] Determine the historical entity information supervector at the current moment, and arrange the historical entity information supervectors;

[0037] The arrangement result is superimposed with the entity information hypervector at the current moment to obtain the calibration action hypervector.

[0038] In one embodiment, the step of calibrating a preset action category based on the calibrated action hypervector and performing gesture recognition on the user's gesture based on the calibrated preset action category includes:

[0039] Obtain the category hypervector corresponding to the preset action category, and determine whether the calibration similarity between the calibration action hypervector and the category hypervector is higher than the preset calibration similarity threshold;

[0040] If so, the category supervector is superimposed and calibrated using the calibration action supervector, and the user's gesture to be recognized is recognized based on the calibrated category supervector.

[0041] In one embodiment, the method further includes:

[0042] When the calibration similarity between the calibration action hypervector and the category hypervector is not higher than a preset calibration similarity threshold, the category hypervector is replaced and calibrated using the calibration action hypervector, and the user's gesture to be recognized is recognized based on the calibrated category hypervector.

[0043] In one embodiment, the step of performing gesture recognition on the user's gesture based on a calibrated preset action category includes:

[0044] The user's gesture to be recognized is collected to obtain gesture data;

[0045] The gesture data to be recognized is subjected to dimensionality upscaling to obtain the hypervector of the action to be recognized;

[0046] Determine the recognition similarity between the action hypervector to be identified and each of the calibrated category hypervectors;

[0047] Gesture recognition is performed on the gesture to be recognized based on the recognition similarity scores.

[0048] In one embodiment, before the step of collecting the user's gesture to be calibrated and obtaining calibration gesture data, the method further includes:

[0049] The system collects preset gestures made by the user according to preset action categories to obtain preset gesture data.

[0050] The preset gesture data is subjected to dimensionality upscaling to obtain the category supervector corresponding to the preset action category.

[0051] In one embodiment, the step of performing a dimensionality-up operation on the preset gesture data to obtain the category hypervector corresponding to the preset action category includes:

[0052] The preset gesture data is preprocessed, and the preprocessing operation includes at least one of data cleaning, data standardization, and data augmentation.

[0053] The preprocessed preset gesture data is subjected to dimensionality upscaling to obtain the category supervector corresponding to the preset action category.

[0054] In one embodiment, after the step of performing gesture recognition on the user's gesture based on a calibrated preset action category, the method further includes:

[0055] Collect new gestures made by the user to obtain new gesture data;

[0056] The newly added gesture data is subjected to a dimensionality increase operation to obtain a new action hypervector;

[0057] The newly added action hypervector updates the calibrated category hypervector, and the updated category hypervector is used to perform gesture recognition on the user's gesture to be recognized.

[0058] In one embodiment, the step of updating the calibrated category hypervector using the newly added action hypervector includes:

[0059] Determine whether the new similarity between the newly added action hypervector and the calibrated category hypervector is higher than a preset new similarity threshold;

[0060] If not, the calibrated category hypervector is updated using the newly added action hypervector.

[0061] Furthermore, to achieve the above objectives, this application also proposes a gesture recognition device, the device comprising:

[0062] The gesture acquisition module is used to acquire the user's gestures to be calibrated and obtain calibration gesture data;

[0063] The data upscaling module is used to perform upscaling operations on the calibration gesture data to obtain calibration action hypervectors;

[0064] The gesture recognition module is used to calibrate a preset action category based on the calibrated action supervector, and to perform gesture recognition on the user's gesture based on the calibrated preset action category.

[0065] In addition, to achieve the above objectives, this application also proposes a storage medium storing a gesture recognition program, which, when executed by a processor, implements the gesture recognition method as described above.

[0066] In addition, to achieve the above objectives, this application also proposes a wearable device, which includes: a memory, a processor, and a gesture recognition program stored in the memory and executable on the processor. When the gesture recognition program is executed by the processor, it implements the gesture recognition method described above.

[0067] This application provides a gesture recognition method, apparatus, storage medium, and wearable device. The method includes: collecting a user's gesture to be calibrated to obtain calibration gesture data; performing a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector; calibrating a preset action category based on the calibration action hypervector; and performing gesture recognition on the user's gesture to be recognized based on the calibrated preset action category.

[0068] This application first collects the user's gestures to be calibrated during calibration, obtaining calibration gesture data. Then, it performs a dimensionality-up operation on this data to obtain a calibration action hypervector. This hypervector is then used to calibrate preset action categories, and the calibrated preset action categories are used to recognize the user's gestures. Compared to existing methods that upload low-dimensional gesture data to a cloud server for calibration, this application first performs a dimensionality-up operation on the calibration gesture data to obtain a calibration action hypervector, allowing for direct local calibration based on the hypervector without uploading to a server, thus reducing calibration time. Attached Figure Description

[0069] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a schematic diagram of the wearable device structure in the hardware operating environment involved in the embodiments of this application;

[0072] Figure 2 This is a flowchart illustrating the first embodiment of the gesture recognition method of this application;

[0073] Figure 3 This is a schematic diagram of the rapid calibration process in the first embodiment of the gesture recognition method of this application;

[0074] Figure 4 This is a schematic diagram of EMG signal data dimensionality upscaling in the first embodiment of the gesture recognition method of this application;

[0075] Figure 5 This is a schematic diagram of IMU signal data dimensionality upscaling in the first embodiment of the gesture recognition method of this application;

[0076] Figure 6 This is a flowchart illustrating the second embodiment of the gesture recognition method of this application;

[0077] Figure 7 This is a flowchart illustrating the third embodiment of the gesture recognition method of this application;

[0078] Figure 8 This is a flowchart of the gesture recognition method in the third embodiment of the present application;

[0079] Figure 9 This is a complete flowchart of the gesture recognition method in the third embodiment of this application;

[0080] Figure 10 This is a flowchart illustrating the fourth embodiment of the gesture recognition method of this application;

[0081] Figure 11 This is the first flowchart used in the fourth embodiment of the gesture recognition method of this application;

[0082] Figure 12 This is a custom-added action flowchart in the fourth embodiment of the gesture recognition method of this application;

[0083] Figure 13 This is a complete flowchart of the gesture recognition method in the fourth embodiment of this application;

[0084] Figure 14 This is a structural block diagram of the first embodiment of the gesture recognition device of this application.

[0085] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0086] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0087] Reference Figure 1 , Figure 1 This is a schematic diagram of the wearable device structure of the hardware operating environment involved in the embodiments of this application.

[0088] like Figure 1 As shown, the wearable device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. In this application, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0089] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on wearable devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0090] like Figure 1 As shown, the memory 1005, which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a gesture recognition program.

[0091] exist Figure 1 In the wearable device shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the user device; the wearable device calls the gesture recognition program stored in the memory 1005 through the processor 1001 and executes the gesture recognition method provided in the embodiments of this application.

[0092] It's important to note that gesture recognition is a crucial method for achieving human-computer interaction. Currently, many methods, such as electromyography (EMG), inertial measurement units (IMUs), photoplethysmography (PPG), and ultrasound, can all achieve gesture recognition. Most existing wearable devices (such as wristband devices) utilize EMG or IMU. This involves collecting EMG data via EMG electrodes on the wristband or IMU data via IMU sensors, then analyzing the data using a gesture recognition model deployed within the device to output the gesture recognition result.

[0093] However, with the widespread adoption of gesture recognition systems, gesture recognition tasks have gradually revealed strong individual differences. That is, different people performing the same action will have different movement habits, speed, force, signal quality, and signal strength. Therefore, gesture recognition models can be quickly calibrated using a small number of user samples.

[0094] However, when calibrating a gesture recognition model, since the collected EMG or IMU data are low-dimensional gesture data, deep learning algorithms are required to calibrate the model. However, deep learning algorithms involve significant computational demands and long training times, and require high precision and performance from hardware. Current wearable devices often lack the computing power to support direct calibration training on internal chips or field-programmable gate arrays (FPGAs). Therefore, the common approach is to upload the collected EMG or IMU data to a cloud server, which then calibrates the gesture recognition model using deep learning algorithms and returns the calibrated model parameters to the wearable device, resulting in a lengthy calibration process.

[0095] Therefore, to address the aforementioned shortcomings, this embodiment provides a gesture recognition method. During calibration, the user's gestures to be calibrated are first collected to obtain calibration gesture data. This data is then upscaled to obtain a calibration action hypervector. Subsequently, a preset action category is calibrated based on this hypervector, and the user's gestures to be recognized are then identified using the calibrated preset action category. Compared to existing methods that upload low-dimensional gesture data to a cloud server for calibration, this embodiment first upscales the calibration gesture data to obtain the calibration action hypervector, allowing for direct local calibration based on the hypervector without uploading to a server, thus reducing calibration time.

[0096] For ease of understanding, the following is combined with Figures 2 to 14The gesture recognition method provided in the embodiments of this application will be described in detail.

[0097] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the gesture recognition method of this application. The first embodiment of the gesture recognition method of this application is presented as follows: Figure 2 As shown, in this embodiment, the specific method includes:

[0098] Step S10: Collect the user's gestures to be calibrated to obtain calibration gesture data.

[0099] It is understood that the method of this embodiment can be applied to wearable devices, wherein the wearable device can be an edge device worn on the user's body, such as a smart wristband or smart bracelet, or other wearable devices. This embodiment does not limit the application of such devices. The execution subject of this method embodiment can be the aforementioned wearable device (hereinafter referred to as the device). This embodiment and the following embodiments will be described using the device as an example.

[0100] It is also understood that the aforementioned device can be equipped with sensors to collect the user's gestures; these could be EMG electrodes, IMU sensors, or both. EMG data can be collected via EMG electrodes, while IMU data can be collected via IMU sensors.

[0101] It should be emphasized that the specific number of EMG electrodes and IMU sensors can be set according to the actual situation, and this embodiment does not impose any restrictions on this. For example, when using EMG electrodes, 8-channel, 16-channel, etc., can be used. When using IMU sensors, it can include one triaxial accelerometer, one triaxial gyroscope, and one triaxial magnetometer. Of course, other numbers of sensors can also be used, and this embodiment does not impose any restrictions on this.

[0102] It should be understood that the aforementioned gestures to be calibrated can be gestures made by the user when the device needs to be calibrated. In actual use, when the user interacts with the device after a long interval or when wearing the device again, and finds that the gesture recognition effect is poor, the device can be controlled to enter the calibration function. After entering the calibration function, the device can guide (e.g., video, audio, etc.) the user to make the relevant gestures to be calibrated. The gestures to be calibrated can be all the gestures in the device's preset gestures, or they can be one or more gestures that have poor effect. The specific gestures can be determined according to the user's selection.

[0103] It should also be noted that the above-mentioned calibration gesture data can be data obtained by the sensor collecting the gesture to be calibrated, which may include the gesture signal corresponding to the gesture to be calibrated and the sensor that collected the gesture signal.

[0104] To ensure the accuracy of the collected results, in this embodiment, the device can also guide the user to make the same gesture to be calibrated a preset number of times (e.g., once or twice), collect the gesture data for each gesture to be calibrated, and then integrate the obtained gesture data to obtain the calibration gesture data.

[0105] For ease of understanding, please refer to Figure 3 , Figure 3 This is a schematic diagram of the rapid calibration process in the first embodiment of the gesture recognition method of this application, as shown below. Figure 3 As shown, in a specific implementation, when a user needs to perform calibration, the aforementioned device can guide the user to make a gesture to be calibrated, and collect the gesture to obtain the corresponding calibration gesture data (i.e., ...). Figure 3 Step a: Specify according to the instructions (specify actions according to the preset action set, under the guidance of the calibration system)).

[0106] Step S20: Perform a dimensionality upscaling operation on the calibration gesture data to obtain the calibration action hypervector.

[0107] It is important to emphasize that, in order to improve data quality, after obtaining the above calibration gesture data, preprocessing operations can be performed on the calibration gesture data (i.e., Figure 3 Step b: Action data preprocessing (data cleaning, data standardization, data augmentation, etc.), wherein the preprocessing operation may include at least one of data cleaning, data standardization and data augmentation, and of course other preprocessing operations are also possible, which are not limited in this embodiment.

[0108] Data cleaning can include, but is not limited to, anomaly detection, data repair, and filtering. Anomaly detection can involve detecting anomalies in the obtained calibration gesture data using a preset anomaly detection threshold and eliminating signals below that threshold. Data repair can involve filling in missing gesture signals in the calibration gesture data using a mean-to-difference method. Filtering can involve bandpass filtering the gesture signals in the calibration gesture data according to recognition requirements.

[0109] When performing data standardization, methods may include, but are not limited to, data rectification and data scaling. Data rectification may involve taking the absolute value of the gesture signal; data scaling may involve normalizing the gesture signal to a floating-point number between 0 and 1, or it may be normalized to a floating-point number within other ranges. This embodiment does not impose any limitations on this.

[0110] When performing signal enhancement, methods such as signal slicing, signal stretching, and adding white noise can be used to enhance the time series data. This is because the number of samples used for calibration and training in this embodiment and subsequent embodiments is relatively small (i.e., only 1 to 2 samples in the above-mentioned 1 or 2 times), and thus signal enhancement can strengthen the calibration and training effects.

[0111] Understandably, after preprocessing, the aforementioned device can perform a dimensionality-up operation on the calibration gesture data to obtain the aforementioned calibration action hypervector. This dimensionality-up operation can be performed by increasing the dimensionality of the calibration gesture data. Since existing methods require complex deep learning algorithms for calibration using low-dimensional gesture data, this embodiment allows the device to perform calibration using simpler calculation methods after increasing the dimensionality of the calibration gesture data. Furthermore, this calibration is completed locally on the device without needing to be uploaded to a cloud server, thus reducing calibration time.

[0112] Furthermore, to enhance the dimensionality, in this embodiment, step S20 includes:

[0113] Step S21: Select the entity used for this gesture recognition. The entity includes at least one of the sensor group or electrode group and the signal amplitude.

[0114] It is important to emphasize that the aforementioned entities can be various components and data used to capture and analyze gesture movements during gesture recognition. For example, in this embodiment, the components used for capture can be sensors (i.e., the aforementioned IMU sensor) or electrodes (i.e., the aforementioned EMG electrodes) used to capture gesture movements. However, since multiple sensors can be set within the aforementioned device in this embodiment, for example, an 8-channel EMG electrode corresponds to 8 EMG electrodes, and an IMU sensor corresponds to 1 triaxial accelerometer, 1 triaxial gyroscope, and 1 triaxial magnetometer, all sensors in the IMU sensor can be referred to as the aforementioned sensor group, and multiple EMGs can be referred to as the aforementioned electrode group. The data used for analysis can be the signal amplitudes that may appear in the gesture signal sequence collected by the sensor group or electrode group. Therefore, the aforementioned entities can include the sensor group, electrode group, and signal amplitudes used for acquisition. However, since gesture acquisition can be achieved by only setting a sensor group or electrode group, the entities in this embodiment can include at least one of the sensor group or electrode group and the signal amplitude.

[0115] In actual use, the above-mentioned device can determine whether the entity to be used is a sensor group or an electrode group based on its own set capture components. When only a sensor group is set, the entity used for this gesture recognition is the sensor group and the signal amplitude. When only an electrode group is set, the entity used for this gesture recognition is the motor group and the signal amplitude. When both are set, the entity used for this gesture recognition is the sensor group, the electrode group, and the signal amplitude.

[0116] Step S22: Determine the entity information of the corresponding entity in the calibration gesture data. The entity information includes entity label information and amplitude signal sequence information.

[0117] The aforementioned entity information can be the information corresponding to the entity used in this gesture recognition, including entity label information and amplitude signal sequence information. This is because during gesture recognition, a gesture signal sequence is collected by each sensor or electrode and analyzed to obtain the gesture recognition result. The entity information involved in the calibration gesture data can include the specific sensor or electrode and the corresponding obtained gesture signal sequence. The aforementioned entity label information can be the label information of each sensor in the sensor group that collected the gesture to be calibrated and / or the label information of each electrode in the electrode group. Since multiple sensors can be set in the device in this embodiment, for example, an 8-channel EMG electrode corresponds to 8 EMG electrodes, and an IMU sensor corresponds to 1 triaxial accelerometer, 1 triaxial gyroscope, and 1 triaxial magnetometer, the device can determine which sensor collected which gesture signal when collecting the gesture to be calibrated, and use the label information corresponding to that sensor as the aforementioned entity label information, thus corresponding to the aforementioned sensor group and electrode group.

[0118] The aforementioned amplitude signal sequence information can be the signal sequence information corresponding to the gesture to be calibrated collected by each sensor in the sensor group and / or each electrode in the electrode group. Since the collected gesture signal is a signal sequence in the form of signal segments, the device can take the sequence segment corresponding to the collected gesture signal as the aforementioned signal sequence when collecting the gesture to be calibrated, and generate the aforementioned amplitude signal sequence information, which corresponds to the aforementioned signal amplitude.

[0119] Step S23: Perform dimensionality upscaling on entity information through spatial and temporal encoding to obtain the hypervector of the action to be calibrated.

[0120] After obtaining entity information, dimensionality enhancement can be performed to increase the dimensionality of the entity information. This allows for direct use of simple calculation methods or calibration in subsequent analysis. Specifically, this can be achieved by spatial encoding and temporal encoding of the entity information.

[0121] Furthermore, in order to perform dimensionality-upgrading operations on entity information, step S23 above includes:

[0122] Step S231: Determine the entity label hypervector corresponding to each entity label information through the first preset mapping relationship.

[0123] The aforementioned entity label hypervector can be a hypervector obtained by mapping the entity label information at each time step to a multidimensional space. In this embodiment, after determining the entity label information to be used and the corresponding amplitude signal sequence information, the device can encode the entity label information corresponding to each sensor in the sensor group and / or each electrode in the electrode group in the spatial dimension, mapping the information to a multidimensional space.

[0124] It should be noted that the aforementioned first preset mapping relationship can be a mapping relationship between each entity label information and the entity label hypervector obtained after spatial dimension encoding. In this embodiment, the device may be equipped with an item memory to store the aforementioned first preset mapping relationship. Therefore, during calibration, the device can directly query the item memory to obtain the entity label hypervector corresponding to each entity label information.

[0125] Step S232: Based on the preset binning rules, the signal amplitude in each amplitude signal sequence information is binned, and the amplitude signal sequence supervector corresponding to each binned signal amplitude is determined through the second preset mapping relationship.

[0126] The supervector of the above amplitude signal sequence can be a supervector obtained after encoding the signal amplitude.

[0127] Understandably, since the signal sequences in the amplitude signal sequence information are entities with continuous similarity relationships, they need to be binned before spatial encoding. Specifically, the signal amplitudes involved can be binned. The aforementioned preset binning rules can be the rules used when binning the signal sequence information during the construction of the aforementioned second preset mapping relationship, which may include the number of binning operations. The aforementioned second preset mapping relationship can be the mapping relationship between the signal amplitudes of each signal sequence information and the amplitude signal sequence hypervectors obtained after spatial dimension encoding. However, since binning has been performed beforehand, the aforementioned second preset mapping relationship can also be the mapping relationship between the binned signal amplitudes and the corresponding amplitude signal sequence hypervectors. Furthermore, the aforementioned second preset mapping relationship can also be stored in the aforementioned project memory.

[0128] Therefore, during actual calibration, the above-mentioned equipment can first perform binning operations on the signal amplitude in each calibration signal sequence according to the preset number of binning operations, divide the signal amplitude into the corresponding bins, and then query the project memory to map the information into a multi-dimensional space, so as to obtain the amplitude signal sequence supervector corresponding to each signal amplitude.

[0129] Step S233: Bind each entity label hypervector to the corresponding amplitude signal sequence hypervector to obtain the bound hypervector of entity information at each time step.

[0130] After obtaining the entity label hypervector and the amplitude signal sequence hypervector, since the sensor or electrode is associated with the acquired signal sequence at any given moment, they can be bound together. The binding result is then used as the binding hypervector for each sensor or electrode in the sensor group at that moment. Similarly, the entity label hypervector and the corresponding amplitude signal sequence hypervector are bound one by one at all moments in this manner, thereby obtaining the binding hypervector for each sensor or electrode at each moment.

[0131] It should be understood that the above binding operation can be achieved through the XOR operation, that is, the output is 1 if the two elements are the same, and -1 if they are different. For example, if it is necessary to bind A and B, the binding operation is denoted as (*), where A = [1, -1, 1, 1, -1, 1, 1], B = [1, -1, -1, 1, -1, 1, -1], and then binding A and B results in A*B = [1, 1, -1, 1, 1, 1, -1].

[0132] Reference Figure 4 , Figure 4 This is a schematic diagram illustrating the dimensionality upgrade of EMG signal data in the first embodiment of the gesture recognition method of this application. (See diagram below.) Figure 4 As shown, if the physical entity used is an EMG electrode (i.e., the electrode group mentioned above) and the signal amplitude is [value missing], the acquired signal sequence is an EMG signal sequence, and the number of EMG electrodes is N (i.e., [value missing]). Figure 4 EMG electrodes 1, 2, ..., N are selected. First, the electrode tag corresponding to each EMG electrode and the acquired EMG signal sequence can be determined. The electrode tags of the N EMG electrodes are denoted as EL1...ELn (i.e., ...). Figure 4 Electrode labels EL1, ..., ELn are used to denot the EMG signal sequences collected by each EMG electrode as ES1...ESn (i.e., ..., ESn). Figure 4 The processed signal sequences ES1, ..., ESN are determined by first identifying the electrode tag EL1 and EMG signal sequence EL1 corresponding to EMG electrode 1, the electrode tag EL2 and EMG signal sequence EL2 corresponding to EMG electrode 2, and so on, until the electrode tag ELn and EMG signal sequence ELn corresponding to EMG electrode N are determined; then, the first preset mapping relationship (i.e., ..., ESN) stored in the project memory is used... Figure 4 After the project memory is used, the amplitude signal sequence hypervectors corresponding to each EMG signal sequence are H_ES1……H_ESn, and the entity tag hypervectors corresponding to each EMG electrode are H_EL1……H_ELn. After the binding operation (i.e. Figure 4After the middle *), the binding supervector at a certain moment is H_E1……H_En, and then H_E1=H_EL1*H_ES1……H_En=H_ELn*H_ESn.

[0133] Similarly, refer to Figure 5 , Figure 5 This is a schematic diagram illustrating the dimensionality upgrade of IMU signal data in the first embodiment of the gesture recognition method of this application. (See diagram below.) Figure 5 As shown, if the entity used is an IMU sensor (i.e., the sensor group mentioned above) and the signal amplitude is specified, the sensor label of the three-axis gyroscope is denoted as ALn (n is 1, 2, or 3) (i.e., ...). Figure 5 The accelerometer is labeled ALn, with each axis denoted as Acc1, Acc2, and Acc3. Similarly, the label for each axis and the acquired signal sequence can be determined first, and then mapped using a first preset mapping relationship (i.e., ...). Figure 5 The corresponding entity tag hypervector H_ALn is obtained from the project memory. The IMU signal sequence acquired by the accelerometer is denoted as ASN (i.e., Figure 5 The processed signal sequence ASN is used to obtain the corresponding amplitude signal sequence supervector H_ASn through the first preset mapping relationship, and then the binding operation is performed (i.e. Figure 5 After the middle *), the binding hypervector at a certain moment is denoted as H_A, then H_A=H_ALn*H_ASn;

[0134] The sensor label of the three-axis gyroscope is denoted as GLn (i.e. Figure 5 The gyroscope is labeled GLn, with each axis denoted as Gyro1, Gyro2, and Gyro3. Similarly, the label for each axis and the acquired signal sequence can be determined first. The corresponding entity label hypervector H_GLn is obtained through a first preset mapping relationship. The IMU signal sequence acquired by this gyroscope is denoted as GSn (i.e., Figure 5 The processed signal sequence GSN is used to obtain the corresponding amplitude signal sequence supervector H_GSn through the first preset mapping relationship. After the binding operation, the bound supervector at a certain moment is denoted as H_G, then H_G=H_GLn*H_GSn;

[0135] The sensor tag of the triaxial magnetometer is denoted as MLn (i.e. Figure 5 The magnetometer is labeled GLn, with each axis denoted as Mag1, Mag2, and Mag3. Similarly, the label for each axis and the acquired signal sequence are determined first. The corresponding entity label hypervector H_MLn is obtained through a first preset mapping relationship. The IMU signal sequence acquired by the magnetometer is denoted as MSn (i.e., ...). Figure 5The processed signal sequence MSn is used to obtain the corresponding amplitude signal sequence supervector H_MSn through the first preset mapping relationship. After the binding operation, the bound supervector at a certain moment is denoted as H_M, then H_M=H_MLn*H_MSn.

[0136] Step S234: Superimpose the binding hypervectors to obtain the entity information hypervectors at each time step, and perform time encoding on the entity information hypervectors at each time step to obtain the calibration action hypervectors.

[0137] It should be noted that since only the binding hypervector of each sensor or electrode was previously obtained, in order to obtain the hypervector of the entire sensor group or electrode group, it is necessary to perform a superposition operation on each binding hypervector. The superposition operation can be an element-wise addition strategy with a standardized operation on the result, where the standardization is set to 1 if the result is greater than or equal to 0, and set to -1 if the result is less than 0. For example, if it is necessary to bind A, B and C, the superposition operation is denoted as (+), where A = [1, -1, 1, 1, -1, 1, 1], B = [1, -1, -1, 1, -1, 1, -1], C = [1, -1, -1, -1, 1, 1, 1]. Then, A, B and C are superimposed to obtain A+B+C = [3, -3, -1, 1, -1, 3, 1]. The standardization operation is denoted as Stand, then Stand(A+B+C) = [1, -1, -1, 1, -1, 1, 1].

[0138] Therefore, in practical applications, based on the above example, in an EMG electrode, since there are N EMG electrodes, the binding hypervectors corresponding to these N EMG electrodes can be superimposed to obtain the superimposed binding hypervector at a certain moment. If the certain moment is t, then the binding hypervector at moment t is used as the aforementioned entity information hypervector, denoted as H_E. t H_E t =H_E1 t +H_E2 t +……H_En t ;

[0139] Similarly, in an IMU sensor, the superposition of the binding hypervectors corresponding to the triaxial accelerometer at time t yields an entity information hypervector denoted as H_A. t H_A t =H_A1 t +H_A2 t +H_A3 t The entity information hypervector is obtained by superimposing the bound hypervectors corresponding to time t of the three-axis gyroscope and denoting it as H_G. t H_G t =H_G1 t +H_G2 t+H_G3 t The entity information hypervector is obtained by superimposing the bound hypervectors corresponding to the triaxial magnetometer at time t, denoted as H_M. t H_M t =H_M1 t +H_M2 t +H_M3 t .

[0140] After the spatial encoding operation described above, the entity information hypervectors at each time step can be obtained. Then, to account for changes in the time dimension, the device can perform temporal encoding on the entity information hypervectors at each time step. The resulting encoding can then be used as the hypervector for the calibration action, completing the dimensionality increase operation (i.e., ...). Figure 3 Step c: Motion data encoding (spatial encoding and temporal encoding of EMG or IMU signals respectively)).

[0141] Furthermore, in order to achieve temporal encoding, in this embodiment, the step of temporally encoding the entity information hypervector at each time point to obtain the calibration action hypervector includes:

[0142] Determine the historical entity information supervectors at the current moment, and arrange the historical entity information supervectors.

[0143] The arrangement result is superimposed with the entity information hypervector at the current moment to obtain the calibration action hypervector.

[0144] It is understood that the above permutation operation can be a cyclic shift to the left or right by a preset number of positions. This embodiment uses a cyclic shift to the right by 1 position for explanation. Let the permutation operation be denoted as (ρ). If A = [1, -1, 1, 1, -1, 1, 1], then one permutation ρ... 1 (A) = [1, 1, -1, 1, 1, -1, 1], arranged twice ρ 2 (A) = [1, 1, 1, -1, 1, 1, -1], arranged three times ρ 3 (A) = [-1, 1, 1, 1, -1, 1, 1], and so on. The above historical entity information hypervector can be the entity information hypervector obtained by arranging and superimposing the previous time step.

[0145] And then continue as Figure 4 and Figure 5As shown, if the signal sequence length is T, then each time step t in T corresponds to an entity information hypervector. For example, for an EMG electrode, at time t=0, since this is the earliest time, there is no historical entity information hypervector. The device can directly use H_E0 as the result after arrangement and superposition at this time. At time t=1, the corresponding historical entity information hypervector is H_E0. The device can then arrange H_E0 once to obtain ρ(H_E0), and then superimpose ρ(H_E0) with the entity information hypervector H_E1 at the current time, resulting in H_E1+ρ(H_E0). Similarly, for t=2, we can obtain H_E2+ρ(H_E1+ρ(H_E0))=H_E2+ρ(H_E1)+ρ 2 (H_E0), and so on, when t=T, we can obtain H_E T +ρ(H_E T-1 )+......+ρ T (H_E0), if the calibration action hypervector is denoted as H_E, then H_E = H_E T +ρ(H_E T-1 )+......+ρ T (H_E0) (i.e.) Figure 4 H_E in T ,ρ(H_E T-1 ), ρ T (H_E0)).

[0146] Similarly, for IMU sensors, if the superposition result corresponding to the three-axis accelerometer is denoted as H_A, the superposition result corresponding to the three-axis gyroscope is denoted as H_G, and the superposition result corresponding to the three-axis magnetometer is denoted as H_M, then the historical entity information supervectors H_A0, H_G0, and H_M0 corresponding to H_A, H_G, and H_M at time t=1 can be determined respectively. These supervectors are then arranged to obtain ρ(H_A0), ρ(H_G0), and ρ(H_M0). Finally, ρ(H_G0) is... Superimposing ρ(H_G0) with the current entity information supervector H_A1 yields H_A1 + ρ(H_A0). Superimposing ρ(H_G0) with the current entity information supervector H_G1 yields H_G1 + ρ(H_G1). Superimposing ρ(H_M0) with the current entity information supervector H_M1 yields H_M1 + ρ(H_M0), and so on. At t = T, H_A = H_A T +ρ(H_A T-1 )+......+ρ T (H_A0)(i.e.) Figure 5 H_A T ,ρ(H_A T-1 ), ρT (H_A0)), H_G=H_G T +ρ(H_G T-1 )+......+ρ T (H_G0) (i.e.) Figure 5 H_G T ,ρ(H_G T-1 ), ρ T (H_G0)), H_M=H_M T +ρ(H_M T-1 )+......+ρ T (H_M0) (i.e.) Figure 5 H_M in T ,ρ(H_M T-1 ), ρ T (H_M0)). If the calibration action hypervector is denoted as H_I, then the superposition results can be superimposed again, i.e., H_I = H_A + H_G + H_M, to obtain the calibration action hypervector.

[0147] Step S30: Calibrate the preset action category based on the calibration action hypervector, and perform gesture recognition on the user's gesture based on the calibrated preset action category.

[0148] It should be understood that the aforementioned preset action categories can be action categories preset by the device for gesture recognition. It should be emphasized that the device may also include an associative memory for storing the aforementioned preset action categories.

[0149] In this embodiment, when performing gesture recognition, the gesture to be recognized made by the user can be collected, and the collected gesture data can be converted into a hypervector form after the above-mentioned dimensionality increase operation. Then, it is compared with each preset action category stored in the associative memory, and the most similar preset action category is taken as the recognition result of the gesture to be recognized.

[0150] Furthermore, during calibration, the aforementioned device can superimpose the obtained calibration action hypervector onto the preset action categories in the Lenovo memory, thereby completing the calibration (i.e., Figure 3 Step d: Calibrate the action hypervector overlay (overlay the calibrated action hypervector onto the corresponding category hypervector in the associative memory). After calibration, the user's gesture can be recognized using the calibrated preset action category.

[0151] This embodiment first collects the user's gestures to be calibrated during calibration, obtaining calibration gesture data. Then, it performs a dimensionality upscaling operation on this data to obtain a calibration action hypervector. Next, it calibrates a preset action category based on this hypervector, and then uses the calibrated preset action category to recognize the user's gestures. Compared to existing methods that upload low-dimensional gesture data to a cloud server for calibration, this embodiment first performs a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector, allowing for direct local calibration based on the hypervector without uploading to a server, thus reducing calibration time.

[0152] Meanwhile, existing methods require uploading to a cloud server for calibration, which necessitates an online environment. In contrast, this embodiment allows calibration to be performed locally on the device, eliminating the need for a network connection and making calibration more convenient.

[0153] In addition, existing deep learning algorithms typically require users to perform multiple actions, such as collecting at least 10 gestures to be calibrated, before calibration can be completed. However, this embodiment encodes and maps each entity into a hyperdimensional vector, and the entity's information is holographically distributed across each unit of the hyperdimensional vector. Therefore, the user only needs to perform fewer actions (e.g., 1 to 2 as mentioned above) to complete the calibration, thus simplifying the calibration process and improving the user experience.

[0154] Reference Figure 6 , Figure 6 This is a flowchart illustrating the second embodiment of the gesture recognition method of this application. Based on the first embodiment described above, the second embodiment of the gesture recognition method of this application is proposed.

[0155] To construct the entity label hypervectors corresponding to each sensor label in the above project's memory, such as Figure 6 As shown, in this embodiment, before step S10, the following steps are also included:

[0156] Step S011: Determine the entity type of each entity;

[0157] Step S012: When the entity type is the first entity type, define the entity according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the entity.

[0158] It should be noted that the aforementioned entity type can be the type of sensor group or electrode group used in the device to collect user gestures, as well as the type of signal amplitude. In this embodiment, the entity type can be divided into a first entity type and a second entity type. Based on the above example, since the entities in this embodiment can be divided into discrete entities and continuous entities, the aforementioned entity type can include discrete entity type and continuous entity type. The entity corresponding to the discrete entity type can include IMU sensors and sparse EMG electrodes (e.g., 8-channel electrodes, 16-channel electrodes, etc.), while the entity corresponding to the continuous entity type can include dense EMG electrodes (e.g., 8*8-channel electrodes, 10*10-channel electrodes, etc.) and signal amplitude.

[0159] Understandably, the first entity type mentioned above can be a discrete entity type. In practical use, when the device uses an IMU sensor or a sparse EMG electrode, the device can define the sensor according to a preset hypervector dimension, randomly generating a hypervector for each sensor or electrode as the entity label hypervector. The preset hypervector dimension can be a dimension defined when converting entity information into a hypervector, and can be any number of dimensions. If the preset hypervector dimension is denoted as D, in this embodiment, D can be greater than or equal to 10000 for better results.

[0160] Furthermore, when the device determines that the sensor group used by the device is an IMU sensor or the electrode group used is a sparse EMG electrode, a D-dimensional hypervector can be randomly generated for each axis of the three-axis accelerometer, three-axis gyroscope and three-axis magnetometer in the IMU sensor as the entity label hypervector. Similarly, a D-dimensional hypervector can be randomly generated for each channel electrode in the sparse EMG electrode as the entity label hypervector.

[0161] Furthermore, since the physical meanings of discrete entities are unrelated, when an entity is of the first entity type, the hypervectors corresponding to the entity information are quasi-orthogonal. In this case, the entity can include sensor groups and electrode groups. Therefore, the aforementioned device can make the hypervectors quasi-orthogonal. Taking Hamming distance similarity as an example, the Hamming distance similarity is 1 when two vectors are orthogonal, and approximately 0.5 when two vectors are quasi-orthogonal. When two vectors are quasi-orthogonal, they can be considered completely dissimilar.

[0162] For example, for an IMU sensor, it can be determined whether each pair of D-dimensional hypervectors generated by each axis is quasi-orthogonal; if so, the hypervector is deemed qualified and stored in the project memory; if not, the corresponding hypervectors are regenerated according to the preset hypervector dimensions and the determination is repeated until they are all quasi-orthogonal. Similarly, for sparse EMG electrodes, it can be determined whether each pair of D-dimensional hypervectors generated by each EMG electrode is quasi-orthogonal; if not, they are randomly generated again.

[0163] It is important to emphasize that in a randomly generated hypervector, each component can be independently and identically distributed, and can take the value of 1 or -1.

[0164] Step S013: Obtain the entity label information corresponding to the entity, and construct a first preset mapping relationship based on the entity label information and the corresponding entity label hypervector.

[0165] After obtaining the entity tag hypervectors corresponding to each sensor in the sensor group and / or each EMG electrode in the electrode group, the device can acquire the entity tag information corresponding to each sensor and / or each EMG electrode, that is, the tag information corresponding to each axis or electrode. After establishing the corresponding mapping relationship, a first preset mapping relationship can be constructed and stored in the aforementioned project memory. During calibration, the entity tag hypervectors can be directly obtained by querying the entity tag information corresponding to each sensor and / or each EMG electrode.

[0166] Furthermore, when the entity type is a continuous entity type, after step S011 above, the following is also included:

[0167] Step S014: When the entity type is the second entity type and the entity is an electrode group, select a sub-entity from the entity.

[0168] It should be understood that the aforementioned second entity type can be a continuous entity type. The aforementioned sub-entity can be the starting electrode in the electrode group used for defining the hypervector. Based on the above example, when the device determines that the entity type is a continuous entity type and the entity is a dense EMG electrode, since dense EMG electrodes are usually in the form of a matrix electrode, such as an X*X electrode sheet attached to the skin, the electrodes adjacent to the vertical axis and the electrodes adjacent to the horizontal axis are similar. Therefore, when the entity is the second entity type, the hypervectors corresponding to the entity information of the entity are similar. At this time, the entity can include the electrode group and the signal amplitude, but the electrode group will be described first. Therefore, for this type of electrode, one electrode can be selected as the aforementioned sub-entity. For example, in this embodiment, the electrode at position (1,1) is used for description.

[0169] Step S015: Define the sub-entity according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the sub-entity.

[0170] After determining that the electrode at position (1,1) is a sub-entity, the device can randomly generate a D-dimensional hypervector for its horizontal axis, denoted as H1, and a D-dimensional hypervector for its vertical axis, denoted as Z1. Then, the hypervectors of the horizontal axis and the hypervectors of the vertical axis are bound together, and the binding result is used as the entity label hypervector corresponding to the sub-entity. That is, for the electrode at position (1,1), the corresponding entity label hypervector is H1*Z1.

[0171] Step S016: Flip the entity label hypervector corresponding to the sub-entity and use the flipped result as the entity label hypervector corresponding to the entity.

[0172] After obtaining the entity tag hypervector H1*Z1, the above device can also use the horizontal axis as a reference to randomly flip the horizontal axis hypervector H1 to generate the horizontal axis hypervector H2 of the adjacent electrode (2,1). Using the vertical axis as a reference, the vertical axis hypervector Z1 is randomly flipped to generate the vertical axis hypervector Z2 of the adjacent electrode (1,2). Then, H2 and Z2 are bound together, and the obtained result is used as the entity tag hypervector corresponding to the electrode, which is H2*Z2.

[0173] Similarly, for subsequent electrodes, their horizontal axis hypervectors are obtained by randomly flipping the horizontal axis hypervector of the previous electrode, which is based on the horizontal axis, and their vertical axis hypervectors are obtained by randomly flipping the vertical axis hypervector of the previous electrode, which is based on the vertical axis. Then, binding is performed to obtain the corresponding entity tag hypervector. For example, the hypervector of the electrode at position (3, 5) is defined as: H3*Z5.

[0174] It is important to emphasize that during random flipping, D / 2 / (X-1) vectors can be flipped, where X is the number of channels and D is the aforementioned preset hypervector dimension. Furthermore, in the randomly generated vectors, each component can be independently and identically distributed, taking values ​​of -1 or 1.

[0175] After obtaining the entity tag hypervectors corresponding to each electrode, the aforementioned device can also obtain the corresponding entity tag information, that is, the electrode tag information corresponding to each electrode. After establishing the corresponding mapping relationship, a first preset mapping relationship can be constructed and stored in the aforementioned project memory. During calibration, the entity tag hypervectors can be directly obtained by querying the entity tag information corresponding to each EMG electrode.

[0176] Furthermore, in order to construct the second preset mapping relationship, in this embodiment, after step S011 above, the following is also included:

[0177] Step S017: When the entity type is the second entity type and the entity is a signal amplitude, the entity is scaled up and down, and the scaled entity is divided into boxes according to the preset boxing rules.

[0178] It should be noted that when the device determines that the entity type is a continuous entity type and the entity is a signal amplitude, the signal amplitude can be scaled down to a floating-point number within a preset scaling range (e.g., 0 to 1). Then, the scaled signal amplitudes can be binned according to a preset binning rule (i.e., the number of binning operations mentioned above). For example, if the number of binning operations is 5, the signal amplitudes can be divided into 0-0.2, 0.2-0.4, 0.4-0.6, 0.6-0.8, and 0.8-1, with each bin corresponding to a range.

[0179] Step S018: Determine the target box based on the results of each sub-box, and define the target box according to the preset vector dimension to obtain the supervector of the amplitude signal sequence corresponding to the target box.

[0180] Understandably, the target box mentioned above can be the starting box when defining the signal amplitude. The device can define a hypervector for each box, but before that, due to the similarity between the boxes, a target box needs to be determined. Continuing with the example above, the device can take the first box as the target box and randomly generate a D-dimensional hypervector as the hypervector of the amplitude signal sequence corresponding to the target box.

[0181] Step S019: Flip the supervector of the amplitude signal sequence corresponding to the target box to obtain the supervector of the amplitude signal sequence corresponding to each sub-box result.

[0182] After obtaining the amplitude signal sequence supervector corresponding to the first box, it can also be randomly flipped D / 2 / (X-1) times, and the flipped result is used as the amplitude signal sequence supervector of the box adjacent to the first box, that is, the amplitude signal sequence supervector of the second box, and so on, until the amplitude signal sequence supervector of all boxes is obtained. This ensures that two adjacent boxes are similar and that the two boxes farthest apart are quasi-orthogonal.

[0183] It should be emphasized that the number of times the boxes are divided can be set according to different tasks, and can generally be 10-50, but this embodiment does not limit this.

[0184] Step S010: Obtain the signal amplitude after each binning result, and construct a second preset mapping relationship based on the signal amplitude after each binning and the corresponding amplitude signal sequence supervector.

[0185] After obtaining the amplitude signal sequence hypervectors corresponding to each bin, a second preset mapping relationship can be established with the corresponding binned signal amplitudes and stored in the project memory. Then, during calibration, the corresponding amplitude signal sequence hypervectors can be directly retrieved based on the binned signal amplitudes.

[0186] In this embodiment, the first and second preset mapping relationships can be pre-built and stored in the local project storage before calibration, eliminating the need for cloud storage. Furthermore, using mapping relationships to implement time and spatial encoding shortens the dimensionality upgrade time and improves calibration efficiency.

[0187] Reference Figure 7 , Figure 7 This is a flowchart illustrating the third embodiment of the gesture recognition method of this application. Based on the above embodiments, the third embodiment of the gesture recognition method of this application is proposed.

[0188] like Figure 7 As shown, considering that during calibration, if the calibration hypervector is directly superimposed onto the preset action category, the similarity between the newly added calibration hypervector and the hypervector corresponding to the preset action category will gradually decrease as the number of superpositions increases, therefore, in this embodiment, the above step S30 includes:

[0189] Step S31: Obtain the category hypervector corresponding to the preset action category, and determine whether the calibration similarity between the calibration action hypervector and the category hypervector is higher than the preset calibration similarity threshold.

[0190] It should be noted that the aforementioned category hypervectors can be hypervectors corresponding to preset action categories. In this embodiment, when the device is used for the first time, the user can be guided to input gesture data corresponding to the relevant preset action categories, and the corresponding category hypervectors can be obtained through the above-mentioned dimensionality increase operation and stored in the associative memory.

[0191] In practical use, after the device obtains the calibration action hypervector, it can determine the similarity between the category hypervector of the preset action category corresponding to the gesture to be calibrated. The similarity comparison strategy can be Hamming distance, cosine similarity or dot product similarity, etc. This embodiment does not limit this.

[0192] After obtaining the corresponding similarity, it is used as the above calibration similarity and compared with the preset calibration similarity threshold.

[0193] Step S32: If yes, then the category supervector is superimposed and calibrated by calibrating the action supervector, and the gesture recognition of the user's gesture is performed based on the calibrated category supervector.

[0194] In specific implementation, when the calibration similarity is higher than the preset calibration similarity threshold, it can be concluded that direct superposition will not reduce the similarity. Therefore, the device can superimpose the calibration action supervector onto the corresponding category supervector to achieve superposition calibration, and then use the calibrated category supervector to recognize the user's gesture in subsequent use.

[0195] Furthermore, after the step of determining whether the calibration similarity between the calibration action hypervector and the category hypervector is higher than a preset calibration similarity threshold, the method further includes:

[0196] Step S33: When the calibration similarity between the calibrated action hypervector and the category hypervector is not higher than the preset calibration similarity threshold, the category hypervector is replaced and calibrated by calibrating the action hypervector, and the user's gesture to be recognized is recognized based on the calibrated category hypervector.

[0197] In the specific implementation, if the calibration similarity is not higher than the preset similarity threshold, it can be said that direct superposition will reduce the similarity. Therefore, the device can directly replace the corresponding category supervector with the calibration action supervector to achieve replacement calibration, and perform gesture recognition on the user's gesture based on the calibrated category supervector.

[0198] Furthermore, in order to achieve gesture recognition, in this embodiment, the step of performing gesture recognition on the user's gesture based on the calibrated preset action category includes:

[0199] Step S331: Collect the user's gesture to be recognized and obtain the gesture data to be recognized;

[0200] Step S332: Perform a dimensionality upscaling operation on the gesture data to be recognized to obtain the hypervector of the action to be recognized.

[0201] It is understandable that the aforementioned hypervector of the action to be recognized can be a hypervector obtained by processing the gesture data to be recognized through the aforementioned time encoding and spatial encoding.

[0202] Reference Figure 8 as well as Figure 9 , Figure 8 This is a flowchart of the gesture recognition method in the third embodiment of this application. Figure 9 This is a complete flowchart of the gesture recognition method in the third embodiment of this application, as follows: Figure 8 and Figure 9 As shown, in actual use, after the device is calibrated, it can collect the user's gestures to be recognized in real time. When the device only has EMG electrodes, it collects data through the EMG electrodes; when it only has an IMU sensor, it collects data through the IMU sensor; and when it has both EMG electrodes and an IMU sensor, it collects data from both. The device then performs motion segment detection on the collected results, extracting the portions containing movement, which are used as the aforementioned gesture data to be recognized (i.e.,...). Figure 8 Step e: Data Acquisition and Activity Segment Detection (The device acquires user EMG and IMU data in real time, and filters actions through activity segment detection) (that is...) Figure 9 (EMG signal, IMU signal).

[0203] After the active segment detection is completed, the gesture data to be recognized can be preprocessed, namely, data cleaning and data standardization. Figure 8 Step f: Action data preprocessing (data cleaning, data standardization, etc.) Figure 9 The data preprocessing is the same as the preprocessing process mentioned above, and will not be described in detail in this embodiment.

[0204] After preprocessing, the gesture data to be recognized can be temporally and spatially encoded, that is, the entity tag information and amplitude signal sequence information can be obtained and retrieved in the project memory. For EMG signals, the corresponding EMG signal encoding method is used (i.e., Figure 4 The corresponding encoding method), for IMU signals, the corresponding encoding method of IMU signals (i.e. Figure 5 The corresponding encoding method) is used to perform dimensionality upscaling on the gesture data to be recognized, thereby obtaining the corresponding hypervector of the action to be recognized (i.e., Figure 8 Step g: Motion data encoding (spatial and temporal encoding of EMG or IMU signals respectively) (i.e.) Figure 9 The system uses the EMG spatial encoder and EMG temporal encoder in the project memory to upscale the EMG signal to obtain H_E, and uses the IMU spatial encoder and IMU temporal encoder in the project memory to upscale the IMU signal to obtain H_I.

[0205] It should be emphasized that this process is consistent with the process of obtaining the calibration action hypervector described above, and will not be elaborated upon in this embodiment. Furthermore, if only EMG electrodes exist, only H_E (i.e., Figure 9 If only the IMU electrode exists, then only H_I (i.e., output1) is obtained. Figure 9 If both H_E and H_I exist, then the superposition of H_E and H_I is taken as the final supervector of the action to be identified, denoted as H_comb (i.e., output2). Figure 9 (output3).

[0206] Step S333: Determine the recognition similarity between the action hypervector to be recognized and each calibrated category hypervector;

[0207] Step S334: Perform gesture recognition based on the recognition similarity scores of each gesture to be recognized.

[0208] After obtaining the hypervector of the action to be recognized, the similarity between the hypervector and the calibrated category hypervector corresponding to each preset action category stored in the associative memory is compared to obtain the corresponding recognition similarity. If all recognition similarities are lower than the preset recognition similarity threshold, the collected gesture to be recognized is determined to be an invalid gesture, and gesture recognition is not performed. If all recognition similarities are not lower than the preset recognition similarity threshold, they are sorted, and the preset action category corresponding to the highest recognition similarity is selected as the gesture recognition result (i.e., Figure 9 (The gesture recognition is achieved by retrieving the action hypervector from the associative memory).

[0209] Reference Figure 10 , Figure 10 This is a flowchart illustrating the fourth embodiment of the gesture recognition method of this application. Based on the above embodiments, the fourth embodiment of the gesture recognition method of this application is proposed.

[0210] To construct category hypervectors corresponding to each preset action category in the association memory, such as Figure 10 As shown, in this embodiment, before step S331, the following steps are also included:

[0211] Step S3301: Collect the preset gestures made by the user according to the preset action category to obtain preset gesture data;

[0212] It should be noted that the preset gestures mentioned above can be gestures that the device allows to recognize. For ease of understanding, please refer to... Figure 11 To explain, Figure 11 This is the first flowchart used in the fourth embodiment of the gesture recognition method of this application, as shown below. Figure 11 As shown, when the device is used by a user for the first time, it can guide the user to perform a specified preset gesture 1 to 2 times according to the preset action category (i.e., Figure 11 Step i: Perform the specified action according to the instructions (perform the specified action according to the preset action set) to obtain the preset gesture data.

[0213] Step S3302: Perform a dimensionality upscaling operation on the preset gesture data to obtain the category supervector corresponding to the preset action category.

[0214] After obtaining the preset gesture data, it can be preprocessed (i.e. Figure 11 Step j: Action data preprocessing (data cleaning, data standardization, data augmentation, etc.), that is, the above step S3302 includes: performing preprocessing operations on preset gesture data, the preprocessing operations including at least one of data cleaning, data standardization and data augmentation; performing dimensionality upscaling operations on the preprocessed preset gesture data to obtain the category supervector corresponding to the preset action category.

[0215] It is understood that the specific preprocessing process is consistent with the preprocessing process mentioned above, and this embodiment will not elaborate on it.

[0216] After preprocessing, the aforementioned device can perform dimensionality upscaling on the preset gesture data. This involves temporal and spatial encoding based on the item memory, extracting entity tag information and amplitude signal sequence information, and querying the item memory. For EMG signals, the corresponding EMG signal encoding method is used; for IMU signals, the corresponding IMU signal encoding method is used. This achieves dimensionality upscaling of the gesture data to be recognized, obtaining the corresponding category supervector (i.e.,...). Figure 11 Step k: Motion data encoding (spatial and temporal encoding of EMG or IMU signals respectively), and storing the obtained category hypervectors with the corresponding preset motion categories in the associative memory (i.e., Figure 11 Step 1: Define and store category hypervectors (define action data as hypervectors for specified action categories and store them in associative memory) (i.e.) Figure 9 (Creating action hypervectors in associative memory).

[0217] It should be emphasized that this process is consistent with the process of obtaining the calibration action hypervector described above, and will not be repeated in this embodiment.

[0218] Furthermore, to enable users to add custom gestures according to their needs and achieve personalized interaction, traditional methods require synchronous modification of network results on the cloud server and device when adding a new gesture, resulting in significant resource consumption. Therefore, in this embodiment, after the step of performing gesture recognition on the user's gesture based on the calibrated preset action category, the method further includes:

[0219] Step S50: Collect the new gestures made by the user to obtain new gesture data;

[0220] Step S60: Perform a dimensionality increase operation on the newly added gesture data to obtain the new action hypervector.

[0221] It should be understood that the aforementioned new gestures can be user-defined gestures. For ease of understanding, please refer to... Figure 12 To explain, Figure 12 As shown in Figure 12, the fourth embodiment of the gesture recognition method of this application provides a flowchart for a custom newly added action. The device also guides the user to perform the custom newly added gesture 2 to 3 times and collects the data (i.e.,...). Figure 12 Step m: Obtain the user's new action (follow the instructions for the user to perform a custom action) and obtain the new gesture data.

[0222] After obtaining new gesture data, it can be preprocessed (i.e. Figure 12Step n: Action data preprocessing (data cleaning, data standardization, data augmentation, etc.). The specific preprocessing process is the same as the preprocessing process mentioned above, and will not be described in detail in this embodiment.

[0223] After preprocessing, the aforementioned device can perform dimensionality upscaling on the newly added gesture data. This involves temporal and spatial encoding based on the project memory, extracting entity tag information and amplitude signal sequence information, and querying the project memory. For EMG signals, the corresponding EMG signal encoding method is used; for IMU signals, the corresponding IMU signal encoding method is used. This process upscales the newly added gesture data to obtain the corresponding new action hypervector (i.e.,...). Figure 12 In step o: motion data encoding (spatial encoding and temporal encoding of EMG signals or IMU signals respectively) to obtain the new motion hypervector.

[0224] Step S70: Update the calibrated category hypervector by adding a new action hypervector, and perform gesture recognition on the user's gesture using the updated category hypervector.

[0225] After obtaining the new action hypervector, the new action hypervector and the corresponding action category can be stored together in the associative memory, thereby updating the total number of calibrated category hypervectors stored in the associative memory. In subsequent gesture recognition, gesture recognition can be achieved based on the updated associative memory.

[0226] Furthermore, to prevent newly added gestures from being too similar to those already stored, in this embodiment, the step of updating the calibrated category hypervector using the newly added action hypervector includes:

[0227] Step S71: Determine whether the new similarity between the newly added action hypervector and the calibrated category hypervector is higher than the preset new similarity threshold;

[0228] Step S72: If not, update the calibrated category hypervector by adding a new action hypervector.

[0229] It should be noted that after obtaining the new action hypervector, the device can determine the similarity between the new action hypervector and each calibrated category hypervector already stored in the associative memory, using this as the new similarity, and compare it with a preset new similarity threshold (i.e., Figure 12Step p: New action validity verification (comparing the similarity of the new action's hypervector with other category hypervectors); if any new similarity exceeds the preset new similarity threshold, it indicates that a similar gesture already exists, and the new gesture is not considered valid, prompting the user to re-enter; if none exceed the threshold, it is considered a valid gesture, and the new action hypervector is stored in the associative memory (i.e., Figure 12 In step q: Store the new action (store the action hypervector that meets the requirements as the new category hypervector)).

[0230] This embodiment implements incremental learning using hypervectors, achieving low resource consumption while meeting the user's need for customized gestures. Furthermore, traditional deep learning incremental learning methods typically only support a fixed number of actions, requiring changes to the network structure and exhibiting poor scalability. This embodiment, however, eliminates the need to consider adding new actions, thus improving scalability.

[0231] Reference Figure 13 , Figure 13 This is a complete flowchart of the fourth embodiment of the gesture recognition method of this application, as follows: Figure 13 As shown, in a specific implementation, the above-mentioned device can initialize the information of each entity tag and the amplitude of each signal before use, thereby obtaining the corresponding entity tag supervector and amplitude signal sequence supervector (i.e., Figure 13 In step r: Hypervector initialization (hypervector initialization of signal amplitude, electrode information, sensor information, etc.); during use, the user's gestures can be collected to obtain the corresponding gesture data (i.e., Figure 13 Step s: Data acquisition (obtaining user dynamic data via EMG electrodes or IMU sensors); and preprocessing the gesture data (i.e. Figure 13 Step t: Data preprocessing (data cleaning, data standardization, data augmentation, etc.); then encoding via project storage (i.e. Figure 13 The intermediate step u: motion data encoding (spatial and temporal encoding of EMG or IMU signals respectively), and finally, based on the encoded hypervectors, functions such as gesture recognition, gesture calibration, and adding new gestures are implemented (i.e. Figure 13 Step v: Execute the function (select to execute gesture recognition, quick calibration, incremental learning function)).

[0232] Furthermore, this application also proposes a storage medium storing a gesture recognition program, which, when executed by a processor, implements the gesture recognition method described above.

[0233] In addition, refer to Figure 14 , Figure 14 This is a structural block diagram of the first embodiment of the gesture recognition device of this application; as shown... Figure 14As shown in the figure, this application also proposes a gesture recognition device, which includes:

[0234] The gesture acquisition module 141 is used to acquire the user's gestures to be calibrated and obtain calibration gesture data;

[0235] The data upsizing module 142 is used to perform upsizing operations on the calibration gesture data to obtain calibration action hypervectors.

[0236] The gesture recognition module 143 is used to calibrate a preset action category based on the calibration action supervector, and to perform gesture recognition on the user's gesture based on the calibrated preset action category.

[0237] This embodiment first collects the user's gestures to be calibrated during calibration, obtaining calibration gesture data. Then, it performs a dimensionality upscaling operation on this data to obtain a calibration action hypervector. Next, it calibrates a preset action category based on this hypervector, and then uses the calibrated preset action category to recognize the user's gestures. Compared to existing methods that upload low-dimensional gesture data to a cloud server for calibration, this embodiment first performs a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector, allowing for direct local calibration based on the hypervector without uploading to a server, thus reducing calibration time.

[0238] As one implementation, the data upsizing module 142 is also used to select the entity used in this gesture recognition, the entity including at least one of the sensor group or electrode group and the signal amplitude; determine the entity information of the corresponding entity in the calibration gesture data, the entity information including entity label information and amplitude signal sequence information; and perform upsizing operation on the entity information through spatial encoding and temporal encoding to obtain the hypervector of the action to be calibrated.

[0239] As one implementation method, the entity type of the entity includes a first entity type and a second entity type; when the entity is of the first entity type, the hypervectors corresponding to the entity information of the entity are quasi-orthogonal, and the entity includes a sensor group and an electrode group; when the entity is of the second entity type, the hypervectors corresponding to the entity information of the entity are similar, and the entity includes an electrode group and a signal amplitude.

[0240] As one implementation method, the data upscaling module 142 is further configured to determine the entity label hypervector corresponding to each entity label information through a first preset mapping relationship; to bin the signal amplitude in each amplitude signal sequence information according to a preset binning rule, and to determine the amplitude signal sequence hypervector corresponding to each binned signal amplitude through a second preset mapping relationship; to bind each entity label hypervector and the corresponding amplitude signal sequence hypervector to obtain the bound hypervector of the entity information at each time step; to superimpose each bound hypervector to obtain the entity information hypervector at each time step, and to perform time encoding on the entity information hypervector at each time step to obtain the calibration action hypervector.

[0241] As one implementation method, the data dimensionality upgrade module 142 is also used to determine the historical entity information supervector at the current moment and arrange the historical entity information supervector; the arrangement result is superimposed with the entity information supervector at the current moment to obtain the calibration action supervector.

[0242] Based on the first embodiment of the gesture recognition device described in this application, a second embodiment of the gesture recognition device of this application is proposed.

[0243] In this embodiment, the gesture acquisition module 141 is also used to determine the entity type of each entity; when the entity type is the first entity type, the entity is defined according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the entity; the entity label information corresponding to the entity is obtained, and a first preset mapping relationship is constructed based on the entity label information and the corresponding entity label hypervector.

[0244] As one implementation, the gesture acquisition module 141 is also used to select a sub-entity from the entity when the entity type is a second entity type and the entity is an electrode group; define the sub-entity according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the sub-entity; flip the entity label hypervector corresponding to the sub-entity and use the flipping result as the entity label hypervector corresponding to the entity.

[0245] As one implementation, the gesture acquisition module 141 is further configured to: scale the entity when the entity type is a second entity type and the entity is a signal amplitude; divide the scaled entity into bins according to a preset binning rule; determine the target bin based on each binning result; define the target bin according to a preset vector dimension; obtain the amplitude signal sequence supervector corresponding to the target bin; flip the amplitude signal sequence supervector corresponding to the target bin to obtain the amplitude signal sequence supervector corresponding to each binning result; obtain the binned signal amplitude corresponding to each binning result; and construct a second preset mapping relationship based on the binned signal amplitude and the corresponding amplitude signal sequence supervector.

[0246] Based on the above embodiments of the gesture recognition device of this application, a third embodiment of the gesture recognition device of this application is proposed.

[0247] In this embodiment, the gesture recognition module 143 is further configured to obtain the category supervector corresponding to the preset action category, and determine whether the calibration similarity between the calibrated action supervector and the category supervector is higher than the preset calibration similarity threshold; if so, the category supervector is superimposed and calibrated by calibrating the action supervector, and the gesture recognition is performed on the user's gesture based on the calibrated category supervector.

[0248] As one implementation, the gesture recognition module 143 is also used to replace and calibrate the category supervector by calibrating the action supervector when the calibration similarity between the calibrated action supervector and the category supervector is not higher than a preset calibration similarity threshold, and to perform gesture recognition on the user's gesture based on the calibrated category supervector.

[0249] As one implementation, the gesture recognition module 143 is also used to collect the gesture to be recognized made by the user to obtain gesture data to be recognized; perform a dimensionality increase operation on the gesture data to be recognized to obtain a hypervector of the action to be recognized; determine the recognition similarity between the hypervector of the action to be recognized and each calibrated category hypervector; and perform gesture recognition on the gesture to be recognized based on each recognition similarity.

[0250] As one implementation, the gesture acquisition module 141 is also used to acquire preset gestures made by the user according to preset action categories to obtain preset gesture data; and to perform dimensionality upscaling operation on the preset gesture data to obtain the category hypervector corresponding to the preset action category.

[0251] As one implementation, the gesture acquisition module 141 is also used to perform preprocessing operations on the preset gesture data, including at least one of data cleaning, data standardization, and data augmentation; and to perform dimensionality upscaling operations on the preprocessed preset gesture data to obtain the category hypervector corresponding to the preset action category.

[0252] As one implementation, the gesture recognition module 143 is also used to collect new gestures made by the user to obtain new gesture data; perform a dimensionality increase operation on the new gesture data to obtain a new action hypervector; update the calibrated category hypervector through the new action hypervector, and perform gesture recognition on the user's gesture to be recognized through the updated category hypervector.

[0253] As one implementation, the gesture recognition module 143 is also used to determine whether the new similarity between the newly added action hypervector and the calibrated category hypervector is higher than a preset new similarity threshold; if so, the calibrated category hypervector is updated by the newly added action hypervector.

[0254] Other embodiments or specific implementations of the gesture recognition device described in this application can be found in the above-described method embodiments, and will not be repeated here.

[0255] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0256] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0257] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0258] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A gesture recognition method, characterized in that, The method includes: The user's gestures to be calibrated are collected to obtain calibration gesture data; The calibration gesture data is subjected to dimensionality upscaling to obtain a calibration action hypervector; The preset action category is calibrated based on the calibrated action hypervector, and the user's gesture to be recognized is recognized based on the calibrated preset action category. The step of performing a dimensionality upscaling operation on the calibration gesture data to obtain a calibration action hypervector includes: Select the entity used in this gesture recognition, the entity including at least one of the sensor group or electrode group and the signal amplitude; Determine the entity information corresponding to the entity in the calibration gesture data, wherein the entity information includes entity label information and amplitude signal sequence information; The entity label hypervector corresponding to each entity label information is determined through a first preset mapping relationship; the signal amplitude in each amplitude signal sequence information is binned according to a preset binning rule, and the amplitude signal sequence hypervector corresponding to each binned signal amplitude is determined through a second preset mapping relationship; each entity label hypervector and the corresponding amplitude signal sequence hypervector are bound together to obtain the bound hypervector of the entity information at each time step; the bound hypervectors are superimposed to obtain the entity information hypervector at each time step, and the entity information hypervector at each time step is time-encoded to obtain the calibration action hypervector; Before the step of collecting the user's gestures to be calibrated and obtaining calibration gesture data, the method further includes: Determine the entity type of each entity; when the entity type is the second entity type and the entity is an electrode group, select a sub-entity from the entity; define the sub-entity according to the preset hypervector dimension to obtain the entity label hypervector corresponding to the sub-entity; flip the entity label hypervector corresponding to the sub-entity and use the flipped result as the entity label hypervector corresponding to the entity.

2. The method of claim 1, wherein, The method further includes: The entity type of the entity includes a first entity type and a second entity type; When the entity is the first entity type, the hypervectors corresponding to the entity information of the entity are quasi-orthogonal, and the entity includes the sensor group and the sparse electrode group; When the entity is the second entity type, the hypervectors corresponding to the entity information of the entity have similarity, and the entity includes a dense electrode group and the signal amplitude.

3. The method of claim 1, wherein, Following the step of determining the entity type of each entity, the method further includes: When the entity type is the first entity type, the entity is defined according to a preset hypervector dimension to obtain the entity label hypervector corresponding to the entity; Obtain the entity label information corresponding to the entity, and construct a first preset mapping relationship based on the entity label information and the corresponding entity label hypervector.

4. The method of claim 3, wherein, Following the step of determining the entity type of each entity, the method further includes: When the entity type is the second entity type and the entity is a signal amplitude, the entity is scaled up and down, and the scaled entity is divided into boxes according to a preset boxing rule. The target box is determined based on the results of each sub-box, and the target box is defined according to the preset vector dimension to obtain the amplitude signal sequence supervector corresponding to the target box; The amplitude signal sequence supervector corresponding to the target bin is flipped to obtain the amplitude signal sequence supervector corresponding to each binning result; Obtain the signal amplitude after binning corresponding to each binning result, and construct a second preset mapping relationship based on the signal amplitude after binning and the corresponding amplitude signal sequence supervector.

5. The method as described in claim 1, characterized in that, The step of temporally encoding the entity information hypervectors at each time point to obtain the calibration action hypervectors includes: Determine the historical entity information supervector at the current moment, and arrange the historical entity information supervectors; The arrangement result is superimposed with the entity information hypervector at the current moment to obtain the calibration action hypervector.

6. The method of any one of claims 1 to 5, wherein, The step of calibrating a preset action category based on the calibrated action hypervector, and then performing gesture recognition on the user's gesture based on the calibrated preset action category, includes: Obtain the category hypervector corresponding to the preset action category, and determine whether the calibration similarity between the calibration action hypervector and the category hypervector is higher than the preset calibration similarity threshold; If so, the category supervector is superimposed and calibrated using the calibration action supervector, and the user's gesture to be recognized is recognized based on the calibrated category supervector.

7. The method of claim 6, wherein, The method further includes: When the calibration similarity between the calibration action hypervector and the category hypervector is not higher than a preset calibration similarity threshold, the category hypervector is replaced and calibrated using the calibration action hypervector, and the user's gesture to be recognized is recognized based on the calibrated category hypervector.

8. The method of claim 7, wherein, The step of performing gesture recognition based on the calibrated preset action category for the user's gesture to be recognized includes: The user's gesture to be recognized is collected to obtain gesture data; The gesture data to be recognized is subjected to dimensionality upscaling to obtain the hypervector of the action to be recognized; Determine the recognition similarity between the action hypervector to be identified and each of the calibrated category hypervectors; Gesture recognition is performed on the gesture to be recognized based on the recognition similarity scores.

9. The method as described in claim 6, characterized in that, Before the step of collecting the user's gestures to be calibrated and obtaining calibration gesture data, the method further includes: The system collects preset gestures made by the user according to preset action categories to obtain preset gesture data. The preset gesture data is subjected to dimensionality upscaling to obtain the category supervector corresponding to the preset action category.

10. The method of claim 9, wherein, The step of performing a dimensionality-up operation on the preset gesture data to obtain the category hypervector corresponding to the preset action category includes: The preset gesture data is preprocessed, and the preprocessing operation includes at least one of data cleaning, data standardization, and data augmentation. The preprocessed preset gesture data is subjected to dimensionality upscaling to obtain the category supervector corresponding to the preset action category.

11. The method of claim 6, wherein, Following the step of recognizing the user's gesture based on the calibrated preset action category, the method further includes: Collect new gestures made by the user to obtain new gesture data; The newly added gesture data is subjected to a dimensionality increase operation to obtain a new action hypervector; The newly added action hypervector updates the calibrated category hypervector, and the updated category hypervector is used to perform gesture recognition on the user's gesture to be recognized.

12. The method as described in claim 11, characterized in that, The step of updating the calibrated category hypervector using the newly added action hypervector includes: Determine whether the new similarity between the newly added action hypervector and the calibrated category hypervector is higher than a preset new similarity threshold; If not, the calibrated category hypervector is updated using the newly added action hypervector.

13. A gesture recognition apparatus, characterized by The device includes: The gesture acquisition module is used to acquire the user's gestures to be calibrated and obtain calibration gesture data; The data upscaling module is used to perform upscaling operations on the calibration gesture data to obtain calibration action hypervectors; The gesture recognition module is used to calibrate a preset action category based on the calibrated action supervector, and to perform gesture recognition on the user's gesture based on the calibrated preset action category. The data upscaling module is further configured to select the entity used in this gesture recognition, the entity including at least one of the sensor group or electrode group and the signal amplitude; determine the entity information corresponding to the entity in the calibration gesture data, the entity information including entity label information and amplitude signal sequence information; determine the entity label hypervector corresponding to each entity label information through a first preset mapping relationship; bin the signal amplitude in each amplitude signal sequence information according to a preset binning rule, and determine the amplitude signal sequence hypervector corresponding to each binned signal amplitude through a second preset mapping relationship; bind each entity label hypervector and the corresponding amplitude signal sequence hypervector to obtain the bound hypervector of the entity information at each time; superimpose each bound hypervector to obtain the entity information hypervector at each time, and perform time encoding on the entity information hypervector at each time to obtain the calibration action hypervector; The gesture acquisition module is also used to determine the entity type of each entity; when the entity type is a second entity type and the entity is an electrode group, a sub-entity is selected from the entity; the sub-entity is defined according to a preset hypervector dimension to obtain the entity label hypervector corresponding to the sub-entity; the entity label hypervector corresponding to the sub-entity is flipped, and the flipping result is used as the entity label hypervector corresponding to the entity.

14. A storage medium, characterized by The storage medium stores a gesture recognition program, which, when executed by a processor, implements the gesture recognition method as described in any one of claims 1 to 12.

15. A wearable device, comprising: The wearable device includes: a memory, a processor, and a gesture recognition program stored in the memory and executable on the processor, wherein the gesture recognition program, when executed by the processor, implements the gesture recognition method as described in any one of claims 1 to 12.

Images