Motion scene recognition methods, devices and electronic equipment

Abstract

This invention discloses a motion scene recognition method, device, and electronic device. By acquiring gravitational acceleration signals and air pressure signals, an initial category of the motion scene is obtained based on the gravitational acceleration signal, a state category is obtained based on the air pressure signal, and the recognition result of the motion scene is determined based on the initial category and the state category. Therefore, through multi-sensor fusion processing, the computational load and power consumption of motion scene recognition can be reduced, while improving device battery life and recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent identification, and in particular to a motion scene identification method and device and electronic equipment. BACKGROUND

[0002] With the development of science and technology, the functions of various electronic devices are becoming more and more complex, and one of them is to identify a motion scene through an electronic device (such as a mobile phone, a wearable device, etc.) carried by a person. Motion scene identification is generally achieved by using sensor signals. However, the calculation amount and power consumption of the prior art are relatively large, and the recognition accuracy of the motion scene is low. SUMMARY

[0003] Therefore, the embodiments of the present application aim to provide a motion scene identification method, device and electronic equipment to reduce the calculation amount and power consumption of a smart wearable device and improve the accuracy and generalization performance of a scene recognition model.

[0004] In a first aspect, the embodiments of the present application provide a motion scene identification method, which comprises:

[0005] obtaining a detection signal, wherein the detection signal comprises a gravity acceleration signal and an air pressure signal;

[0006] obtaining an initial category of a motion scene according to the gravity acceleration signal;

[0007] obtaining a state category according to the air pressure signal; and

[0008] determining a recognition result of the motion scene according to the initial category and the state category.

[0009] In some embodiments, the obtaining of the detection signal comprises:

[0010] receiving a gravity acceleration signal sent by a three-axis gravity acceleration sensor; and

[0011] receiving an air pressure signal sent by an air pressure gauge.

[0012] In some embodiments, the obtaining of the initial category of the motion scene according to the gravity acceleration signal comprises:

[0013] sampling the gravity acceleration signal through the first window according to a predetermined first length, a first step and a first sampling frequency to obtain the first sampling signal; and

[0014] obtaining the initial category according to the first sampling signal through a pre-trained classification model.

[0015] In some embodiments, the pre-trained classification model comprises an attention layer, a weighting layer, a first convolutional layer, a second convolutional layer, a merging layer, a third convolutional layer, and a fully connected layer;

[0016] wherein the obtaining the initial category from the first sampled signal by the pre-trained classification model comprises:

[0017] obtaining a first intermediate vector from the first sampled signal by the attention layer;

[0018] obtaining a second intermediate vector from the first sampled signal and the first intermediate vector by the weighting layer;

[0019] generating a third intermediate vector from the second intermediate vector by the first convolutional layer;

[0020] generating a fourth intermediate vector from the third intermediate vector by the second convolutional layer;

[0021] generating a fifth intermediate vector from the third intermediate vector and the fourth intermediate vector by the merging layer;

[0022] generating a sixth intermediate vector from the fifth intermediate vector by the third convolutional layer; and

[0023] obtaining the initial category from the sixth intermediate vector by the fully connected layer.

[0024] In some embodiments, the obtaining the state category from the air pressure signal comprises:

[0025] sampling the air pressure signal by the second window to obtain the second sampled signal according to a predetermined second length, a second step, and a second sampling frequency; and

[0026] obtaining the state category from the second sampled signal.

[0027] In some embodiments, the obtaining the state category from the second sampled signal comprises:

[0028] obtaining a differential signal by differentiating the second sampled signal by a first sub-window and a second sub-window; and

[0029] comparing the differential signal with a predetermined threshold to obtain the state category.

[0030] In some embodiments, the initial category comprises at least one of walking, running, cycling, driving, elliptical, and a first category, and the state category comprises at least one of going upstairs, going downstairs, and a second category;

[0031] The initial category and the state category determine the identification result of the motion scene, specifically:

[0032] In response to the initial category being at least one of walking and running, and the state category being at least one of ascending and descending, the initial category and the state category are fused to determine the motion scene identification result.

[0033] In some embodiments, the fusing of the initial category and the state category to determine the motion scene identification result includes at least one of:

[0034] In response to the initial category being walking, and the state category being ascending, it is determined that the identification result of the motion scene is walking down stairs.

[0035] In response to the initial category being running, and the state category being ascending, it is determined that the identification result of the motion scene is running down stairs.

[0036] In response to the initial category being walking, and the state category being descending, it is determined that the identification result of the motion scene is walking up stairs.

[0037] In response to the initial category being running, and the state category being descending, it is determined that the identification result of the motion scene is running up stairs.

[0038] In response to the initial category being at least one of the first category, and the state category being ascending, it is determined that the identification result of the motion scene is elevator down.

[0039] In response to the initial category being at least one of the first category, and the state category being descending, it is determined that the identification result of the motion scene is elevator up.

[0040] In a second aspect, an embodiment of the present application provides a motion scene identification device, the device comprising:

[0041] A detection signal acquisition unit is configured to acquire a detection signal, the detection signal including a gravity acceleration signal and an air pressure signal.

[0042] An initial category acquisition unit is configured to acquire an initial category of a motion scene according to the gravity acceleration signal.

[0043] A state category acquisition unit is configured to acquire a state category according to the air pressure signal.

[0044] An identification result determination unit is configured to determine an identification result of the motion scene according to the initial category and the state category.

[0045] In a third aspect, an electronic device is provided, the electronic device comprising:

[0046] a triaxial gravity acceleration sensor configured to obtain a gravity acceleration signal;

[0047] a barometer configured to obtain a barometric pressure signal; and

[0048] a memory and a processor, the memory being configured to store one or more computer program instructions, wherein the one or more computer program instructions are executable by the processor to implement the method according to the first aspect.

[0049] The technical solution of the embodiments of the present application obtains a gravity acceleration signal and a barometric pressure signal, obtains an initial category of a motion scene according to the gravity acceleration signal, obtains a state category according to the barometric pressure signal, and determines a recognition result of the motion scene according to the initial category and the state category. Thus, through multi-sensor fusion processing, the computational complexity and power consumption of motion scene recognition can be reduced, and the device battery life and recognition accuracy can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] The above and other objects, features and advantages of the present application will become more apparent from the following description of embodiments of the present application taken in conjunction with the accompanying drawings, in which:

[0051] Figure 1 is a schematic diagram of an electronic device according to an embodiment of the present application;

[0052] Figure 2 is a flowchart of a motion scene recognition method according to an embodiment of the present application;

[0053] Figure 3 is a flowchart of obtaining an initial category according to an embodiment of the present application;

[0054] Figure 4 is a waveform diagram of a gravity acceleration signal according to an embodiment of the present application;

[0055] Figure 5 is a schematic diagram of a classification model according to an embodiment of the present application;

[0056] Figure 6 is a flowchart of a classification model obtaining an initial category according to an embodiment of the present application;

[0057] Figure 7 is a flowchart of obtaining a state category according to an embodiment of the present application;

[0058] Figure 8 is a waveform diagram of a barometric pressure signal according to an embodiment of the present application;

[0059] Figure 9 is a schematic diagram of a motion scene recognition apparatus according to an embodiment of the present application;

[0060] Figure 10 is a schematic diagram of performance test of the gravity acceleration signal of an embodiment of the present application;

[0061] Figure 11 is a schematic diagram of performance test of the air pressure signal of an embodiment of the present application. DETAILED DESCRIPTION

[0062] The present application is described in detail below based on embodiments, but the present application is not limited to only these embodiments. In the following detailed description of the present application, some specific details are described in detail. The present application can also be fully understood without the description of these details by those skilled in the art. In order to avoid confusion of the essence of the present application, well-known methods, processes, flows, elements and circuits are not described in detail.

[0063] In addition, those of ordinary skill in the art will understand that the drawings provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0064] Unless the context clearly requires otherwise, throughout the description, the words "comprise", "comprising", and the like are to be construed in an inclusive sense as opposed to an exclusive or exhaustive sense; that is to say, in the sense of "including, but not limited to".

[0065] In the description of the present application, it should be understood that the terms "first", "second", etc. are only for the purpose of description and should not be understood as indicating or implying relative importance. In addition, in the description of the present application, unless otherwise stated, the meaning of "multiple" is two or more.

[0066] Motion scene recognition is a common function of electronic devices such as mobile phones, wearable devices (such as bracelets, smart watches, etc.). Motion scene recognition is generally achieved by taking the signals collected by sensors as input signals.

[0067] The input signal form of the motion scene recognition method of the prior art is mainly realized by three ways: only inputting three-axis gravity acceleration signals, simultaneously inputting three-axis gravity acceleration signals and six-axis inertial navigation signals of a gyroscope, and inputting three-axis gravity acceleration signals and barometer signals.

[0068] Among the three types of signal input forms used in motion scene recognition, the method of only inputting three-axis gravity acceleration signals is simple and easy to operate, but it is difficult to distinguish more categories of motion scenes and has low recognition accuracy.

[0069] Meanwhile, inputting the three-axis gravity acceleration signal and the six-axis inertial navigation signal of the gyroscope in the form, although it can identify more types of motion scene, but also leads to the increase of the amount of data processed by the algorithm, and the opening of the gyroscope generally consumes several times the power of the three-axis gravity accelerometer, resulting in a decrease in the endurance time of the device.

[0070] Inputting the three-axis gravity acceleration signal and the barometer signal can identify more types of motion scene than inputting only the three-axis acceleration signal, while requiring fewer resources and lower power consumption than inputting the three-axis gravity acceleration signal and the six-axis inertial navigation signal of the gyroscope. The prior art based on the use of the three-axis gravity acceleration signal and the barometer signal mainly includes three methods for processing signals: traditional signal processing, machine learning method, and deep learning method.

[0071] Among them, the method based on traditional signal processing, such as time-frequency feature double-threshold method, autocorrelation function method, etc., uses the calculated threshold value as the decision condition for pattern recognition; the second is the machine learning method, such as support vector machine method, artificial neural network method, etc.; the third is the deep learning method, such as convolutional neural network, long short-term memory network, etc.

[0072] Among the three methods of using three-axis gravity acceleration and barometer signals for motion scene recognition in smart watches, the method based on traditional signal processing is simple to design and easy to implement, but the effect of separating different types of motion scenes is not ideal in low signal-to-noise ratio conditions; the machine learning method usually has a relatively high accuracy, but has the disadvantages of relying on prior knowledge, complex operation and large amount of calculation; deep learning is a common method for motion scene recognition in smart watches, which has automatic feature engineering and does not need to rely on too much prior knowledge, and can achieve an algorithm with high recognition accuracy through a data-driven method.

[0073] Usually, when using deep learning to identify motion scenes in smart watches, the first step is to use a sliding window method to intercept the sensor signal sequence input, so when using the three-axis gravity acceleration signal and the barometer signal as the input for motion scene recognition in smart watches, it is required that the three-axis gravity acceleration signal and the barometer signal have the same sampling frequency. The three-axis gravity acceleration signal can be accurately identified at a higher sampling frequency, so the sampling frequency of the barometer signal needs to be consistent with that of the three-axis gravity acceleration signal, while the barometer signal only needs to be at a lower sampling frequency to meet the accuracy requirement, so keeping the same sampling frequency as the three-axis gravity acceleration signal will result in high computational load and power consumption.

[0074] Therefore, embodiments of the present invention provide an electronic device and a motion scene recognition method to reduce the high computational load and power consumption caused by inconsistent or excessively high sampling frequencies when the sliding window method truncates the sensor signal sequence input. Furthermore, it improves the device's battery life and recognition accuracy by adaptively weighting the axes that affect the motion scene recognition and classification of the three-axis gravitational acceleration, and by using a temporal convolutional network to extract the spatiotemporal features contained in the three-axis gravitational acceleration signal.

[0075] Figure 1 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Figure 1 In the illustrated embodiment, the electronic device includes a triaxial accelerometer 11, a barometer 12, a memory 13, and a processor 14. The triaxial accelerometer 11 is used to acquire gravitational acceleration signals. The barometer 12 is used to acquire air pressure signals. The memory 13 is used to store one or more computer program instructions, which are executed by the processor 14 to implement the motion scene recognition method of this embodiment.

[0076] The triaxial accelerometer 11 is used to acquire triaxial gravity acceleration signals. Specifically, the triaxial accelerometer is an accelerometer used to measure spatial acceleration. It primarily decomposes spatial acceleration along the X, Y, and Z axes; technically, three single axes can be combined into a triaxial sensor. For example, assuming that in horizontal walking motion, the vertical and forward accelerations exhibit periodic changes, during the foot-retraction movement, the vertical acceleration increases positively due to the upward center of gravity and single-foot contact with the ground. Continuing forward, the center of gravity shifts downward, and both feet touch the ground, resulting in the opposite acceleration. Horizontal acceleration decreases during foot retraction and increases during stepping. Furthermore, in walking motion with uncertain directions, the vertical and forward accelerations roughly follow a sine curve with time, exhibiting a peak at a certain point. The vertical acceleration shows the greatest variation. By monitoring and calculating the peak value of the trajectory and determining the acceleration threshold, the number of steps can be calculated in real time, and the user's walking distance can be further estimated. By calculating the vector lengths of the three accelerations, a sinusoidal trajectory of walking motion is obtained. Peak detection is then performed, recording the previous vector length and acceleration direction. Changes in vector length determine the current acceleration direction, which is compared to the previously saved direction. If the directions are opposite, indicating the peak has just passed, step counting begins; otherwise, the motion is discarded. Accumulating the number of peaks yields the walking stride pattern. Furthermore, when simulating human walking by briefly and rapidly shaking the device, interference data, if not filtered out, can affect the accuracy of step counting. However, this interference can be filtered out by comparing the detected data with thresholds and using step frequency analysis.

[0077] At present, the main types of three-axis gravity acceleration sensor are: piezoresistive, piezoelectric and capacitive. The acceleration of the measured object is converted into the change of resistance, voltage and capacitance of the sensor unit, and then the change value of the sensor unit is converted into a voltage value through a conversion circuit. Then, the signal is amplified and filtered, and the analog quantity is processed into a suitable stable output signal. At this time, the output signal is a voltage value. Finally, the analog-to-digital converter is used to convert the signal into a digital signal.

[0078] The barometer 12 is used to obtain the air pressure signal. The main sensing element of the barometer is a thin film sensitive to the strength of the gas in the pressure sensor and a needle control. In terms of circuit, it is connected with a flexible resistor. When the pressure intensity of the gas around the measured environment decreases or increases, the thin film deforms to drive the needle, and the resistance of the resistor changes. The resistance also changes, causing the capacitance to change and generating an electric signal. Finally, the analog-to-digital converter is used to convert the signal into a digital signal.

[0079] For example, assume that when the pressure of the measured gas decreases or increases, the thin film deforms and the resistance of the resistor changes. At this time, a 0-5V signal voltage is obtained through the sensor element, which is converted by the single-chip microcomputer and then received by the data collector. Then, the data collector transmits the result to the processor in a suitable form.

[0080] The processor 14 is connected with the three-axis gravity acceleration sensor 11 and the barometer 12 to receive the gravity acceleration signal sent by the three-axis gravity acceleration sensor 11 and the air pressure signal sent by the barometer 12. The memory 13 is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor 14 to realize the motion scene recognition method of the embodiment of the application.

[0081] Specifically, Figure 2 is a flowchart of the motion scene recognition method of the embodiment of the application. As Figure 2 shown, the motion scene recognition method of the embodiment of the application includes the following steps:

[0082] Step S100, obtaining a detection signal, the detection signal including a gravity acceleration signal and an air pressure signal.

[0083] In this embodiment, obtaining the detection signal specifically refers to receiving the gravity acceleration signal sent by the three-axis gravity acceleration sensor, and receiving the air pressure signal sent by the barometer.

[0084] The gravity acceleration signal is an acceleration data generated by measuring the displacement of the human body in a situation where the motion direction is not known in advance, and is converted into the change value of the resistance, voltage and capacitance of the sensing unit, and then converted into a voltage value by the single-chip microcomputer and converted into a digital signal by the digital-to-analog conversion, that is, the gravity acceleration signal. The air pressure signal is an electric signal generated by measuring the change of the resistance caused by the change of the air pressure of the surrounding environment, resulting in the change of the capacitance, and then processed by the single-chip microcomputer to obtain the air pressure signal.

[0085] Step S200, obtaining an initial category of a motion scene according to the gravity acceleration signal.

[0086] In this embodiment, after obtaining the detection signal, an initial category of a motion scene is obtained according to the gravity acceleration signal in the detection signal.

[0087] Specifically, Figure 3 is a flowchart of obtaining an initial category according to an embodiment of the present application. As shown in Figure 3 obtaining an initial category of a motion scene according to the gravity acceleration signal includes the following steps:

[0088] Step S210, sampling the gravity acceleration signal through the first window according to a predetermined first length, first step length and first sampling frequency to obtain the first sampling signal.

[0089] The first length refers to the length of the first window, that is, the length of the time period intercepted each time sampling. The first step length refers to the step length of the first window, that is, the moving time of adjacent two samplings. The first sampling frequency refers to the frequency of obtaining the sampling signal from the received gravity acceleration signal, that is, the sampling points obtained per unit time.

[0090] Figure 4 is a waveform diagram of the gravity acceleration signal according to an embodiment of the present application. In Figure 4 the embodiment shown, the signal change curves of the gravity acceleration sensor in the X-axis, Y-axis and Z-axis are shown, wherein the abscissa is time t, the unit is s (second), and the ordinate is signal value.

[0091] Suppose the first length of the first window is 2 seconds, the first step length is 1 second, and the first sampling frequency is 25 Hz, then:

[0092] In the first sampling period, since the first length is 2s, the waveform of 0s-2s is intercepted in the waveform diagram, and since the first sampling frequency is 25Hz, 25 sampling points are obtained in the time period of 0s-1s and 1s-2s, respectively. At the same time, since the signals of three axes are sampled at the same time, the first sampling signal corresponding to the gravitational acceleration can be expressed as a vector of T1*F1*N1, where T1 is the first length, F1 is the first sampling frequency, and N1 is the number of axes of the gravitational acceleration signal. When the first length is 2s, the first step length is 1s, and the first sampling frequency is 25Hz, the first sampling signal can be expressed as a vector of 2*25*3.

[0093] In the second sampling period, since the first step length is 1s, it is necessary to move 1s, intercept the waveform of 1s-3s in the waveform diagram, and then obtain the first sampling signal of the second sampling period based on the same manner as the first sampling period. In this way, as time goes on, the count period contained in the sliding window moves forward by 1s, but always contains a time period of 2s. Thus, the first sampling signal can be periodically obtained through the first window.

[0094] It should be understood that the embodiments of the present application do not limit the parameters of the first window, and the first length, the first step length and the first sampling frequency can be set according to the actual application scenario.

[0095] In step S220, the initial category is obtained from the first sampling signal through a pre-trained classification model.

[0096] In some embodiments, the classification model can be implemented by a CAM-TCN-net (Class Activation Mapping Temporal Convolutional Network). The CAM-TCN-net is composed of a CAM (Class Activation Mapping) module and a TCN (Temporal Convolutional Network) module, and is used for action recognition tasks. The CAM module generates a class activation map for each category through a channel attention mechanism to learn the significant areas that distinguish different categories. The TCN module uses a one-dimensional convolutional network to learn time sequence features to model long-term action processes. The combination of the two modules can fully utilize spatial information and time sequence information to improve the representation and recognition ability of the model for complex actions.

[0097] Figure 5 is a schematic diagram of the classification model of the embodiments of the present application. The classification model includes an attention layer L1, a weighting layer L2, a first convolutional layer L3, a second convolutional layer L4, a merging layer L5, a third convolutional layer L6 and a fully connected layer L7.

[0098] Figure 6 is a flowchart of the process of obtaining an initial category by the classification model of the embodiment of the present application. In combination with the structure of the classification model shown in Figure 5 , obtaining the initial category by the pre-trained classification model according to the first sampling signal comprises the following steps:

[0099] Step S221, obtaining a first intermediate vector according to the first sampling signal by the attention layer.

[0100] wherein the attention layer L1 and the weighting layer L2 constitute the CAM module of the embodiment of the present application. The attention layer L1 comprises a pooling layer and a full connection layer, wherein the pooling layer is used to receive the first sampling signal V S1 and perform average pooling processing on the first sampling signal V S1 to obtain a compressed vector, and the full connection layer is used to perform equal-dimension full connection processing on the compressed vector to obtain the first intermediate vector V M1 . As described above, the first sampling signal can be expressed as a vector of T1*F1*N1, wherein T1 is the first length, F1 is the first sampling frequency, and N1 is the number of axes of the gravity acceleration signal. The first sampling signal is subjected to average pooling processing by the pooling layer to obtain a vector of 1*1*N1, that is, the pooling layer performs compression processing on the first sampling signal. The average pooling is one of the pooling operations and is a method of reducing sampling, which can reduce the data amount and extract the main features. Its working principle is that the first sampling signal is divided into multiple rectangular pooling windows, and the average value of all elements in each pooling window is calculated as the output of the window. The compressed vector is taken as the input of the full connection layer, and the first intermediate vector is output by the full connection layer. Specifically, assuming that M first sampling signals are obtained, M first intermediate vectors are obtained after being processed by the pooling layer, each first intermediate vector is taken as a hidden state, and a score (scalar) of each hidden state is obtained by a scoring function. All scores are run through a softmax (normalized exponential function) layer so that the sum of the softmax scores (scalar) is 1. These scores represent the distribution of attention. Each hidden state is multiplied by its softmax score (scalar) to obtain an alignment vector. The alignment vectors are aggregated to obtain a context vector. The context vector is input into a decoder, and the first intermediate vector V M1 is output by the decoder.

[0101] wherein the full connection layer is used to map the information feature value to the specimen space for high purification, and plays a role of a classifier in the convolutional neural network.

[0102] In one embodiment, the fully connected layer can be composed of two parts, including a linear part and a nonlinear part.

[0103] Where the linear part is linear weighted summation, for example, for an input vector h = [h0, h1,..., hn] T Where h = [h0, h1,..., hn] T is the transpose of [h0, h1,..., hn], the output vector of the linear part is i = [i0, i1,..., in] T Where i = [i0, i1,..., in] T is the transpose of [i0, i1,..., in], the parameters of the linear part can be imagined as an m*n matrix W, plus a bias term k = [k0, k1,..., kn] T Where k = [k0, k1,..., kn] T is the transpose of [k0, k1,..., kn], so there is W*h+k=i. It is equivalent to analyzing and summarizing the input signal from multiple angles to get multiple output results.

[0104] Where the nonlinear part needs to do a function conversion, which is to normalize the data. For example, assuming there is a 2-layer fully connected neural network, where there is no nonlinear layer, then for the first layer:

[0105] W 0 *h 0 +k 0 =i 1

[0106] For the second layer:

[0107] W 1 *i+bk 1 =i 2

[0108] The two formulas are combined:

[0109] W 1 *(W 0 *h 0 +k 0 )+k 1 =i 2

[0110] W 1 *W 0 *h 0 +(W 1 *k 0 +k 1 )+k 1 =zi 2

[0111] So as long as W 0 = W 1 * W 0 , k 0′ = W 1 * k 0 + k 1 , the previous two-layer neural network can be represented by a one-layer neural network.

[0112] It should be noted that the algorithm of the full connection layer described above is only an example provided by the embodiment of the present application, and the algorithm of the full connection layer is not limited by the embodiment of the present application, which can be implemented based on various existing algorithms.

[0113] The initial category includes at least one of walking, running, cycling, driving, elliptical machine and the first category.

[0114] In step S222, a second intermediate vector is obtained by the weighting layer according to the first sampling signal and the first intermediate vector.

[0115] The weighting layer L2 is configured to obtain a second intermediate vector V S1 from the first sampling signal V M1 and the first intermediate vector V M2 . Specifically, the weighting layer L2 is configured to perform weighted summation on N1 in the first intermediate vector V M1 and the first sampling signal V S1 to obtain the second intermediate vector V M2 .

[0116] In step S223, a third intermediate vector is generated by the first convolution layer according to the second intermediate vector.

[0117] The first convolution layer L3, the second convolution layer L4, the merging layer L5, the third convolution layer L6 and the full connection layer L7 constitute a TCN module of the embodiment of the present application.

[0118] The first convolution layer L3 is configured to generate a third intermediate vector V M2 from the second intermediate vector V M3 . In some embodiments, the first convolution layer L3 can be implemented by a temporal convolutional network (TCN). The temporal convolutional network includes a repeated stack of multi-layer network and a one-dimensional convolution kernel, wherein the second intermediate vector V M2 is processed by the one-dimensional convolution kernel.performing convolution operation, capturing local features in time series, expanding receptive field by repeatedly stacking multi-layer network, learning time series features in different time scales, repeatedly stacking multi-layer network includes multiple continuous sub-networks, each sub-network performs the same principle, the input of the previous sub-network is the output of the next sub-network, the input of the first sub-network is the second intermediate vector V M2 The output of the last sub-network is connected in residual with the local features extracted by the one-dimensional convolution kernel to obtain the third intermediate vector V M3 For each sub-network, including dilated causal convolution (Dilated GausalConv), weight normalization (WeightNorm), activation function (ReLU) and random dropout (Dropout). Dilated causal convolution (Dilated Causal Convolution) is a convolution operation for processing time series, which combines dilated convolution and causal convolution, which can aggregate a wider range of past information without using future information. Dilated convolution is to adjust the distance between elements in the convolution kernel, which can expand the receptive field and obtain more extensive information. Causal convolution only uses information at the current and previous time, and does not use information at the next time, which ensures that the output only depends on the current and past, and avoids information leakage. By stacking multiple dilated causal convolutions with gradually increasing dilation rates, the receptive field can be exponentially expanded, which can effectively learn the long-term dependencies in time series. Dilated causal convolution significantly enhances the model's ability to learn time series data by combining dilated and causal mechanisms, and plays a key role in processing time-dependent tasks. Weight normalization is to normalize the weight value, which has small time overhead, fast operation speed, and introduces less noise. Activation function can make the network training faster, increase the nonlinearity of the network, improve the expression ability of the model, prevent gradient disappearance, and make the network have sparsity, etc. Dropout is to temporarily discard the neural network unit from the network according to a certain probability during the training process of the deep learning network, which can prevent overfitting and improve the operation speed of the model.

[0119] In step S224, a fourth intermediate vector is generated from the third intermediate vector by the second convolution layer.

[0120] The second convolution layer L4 is used to generate a fourth intermediate vector V M3 from the third intermediate vector V M4 In some embodiments, the second convolution layer L4 can be implemented by a temporal convolutional network (TCN), and the specific principle is similar to that of the first convolution layer L3, except that the input and output information of the first convolution layer are different, which will not be described here.

[0121] Step S225, generating a fifth intermediate vector according to the third intermediate vector and the fourth intermediate vector through the merging layer.

[0122] The merging layer L5 is configured to generate a fifth intermediate vector V M3 according to the third intermediate vector V M4 and the fourth intermediate vector V M5 . Specifically, the merging layer L5 is configured to perform a summation operation on the third intermediate vector V M3 and the fourth intermediate vector V M4 to obtain the fifth intermediate vector V M5 .

[0123] Step S226, generating a sixth intermediate vector according to the fifth intermediate vector through the third convolution layer.

[0124] The third convolution layer L6 is configured to generate a sixth intermediate vector V M5 according to the fifth intermediate vector V M6 . In some embodiments, the third convolution layer L6 can be implemented by a temporal convolutional network (TCN), and the specific principle is similar to that of the first convolution layer L3, except that the input and output information of the first convolution layer are different, which will not be described herein.

[0125] Step S227, obtaining the initial category according to the sixth intermediate vector through the fully connected layer.

[0126] The fully connected layer L7 is configured to generate an initial category V M6 according to the sixth intermediate vector V C1 . Specifically, the fully connected layer is a commonly used layer in neural networks, which can recombine the features in the sixth intermediate vector V M6 to obtain the initial category.

[0127] In this embodiment, the initial category includes at least one of walking, running, cycling, driving, elliptical machine and a first category, wherein the first category represents a set of other categories except walking, running, cycling, driving and elliptical machine.

[0128] Step S300, obtaining a state category according to the air pressure signal.

[0129] In this embodiment, after obtaining the detection signal, a state category of the motion scene is obtained according to the air pressure signal in the detection signal.

[0130] Figure 7 is a flowchart of obtaining a state category according to the air pressure signal. As shown in FIG. 7, the flowchart includes the following steps.Figure 7 As shown, acquiring the state category of the motion scene according to the air pressure signal comprises the following steps:

[0131] In step S310, the air pressure signal is sampled through the second window according to a predetermined second length, a second step length and a second sampling frequency to obtain the second sampling signal.

[0132] The second length refers to the length of the second window, that is, the length of the time period intercepted each time sampling. The second step length refers to the step length of the second window, that is, the moving time of adjacent two samplings. The second sampling frequency refers to the frequency of obtaining the sampling signal from the received air pressure signal, that is, the sampling points obtained per unit time.

[0133] Figure 8 is a waveform diagram of the air pressure signal of an embodiment of the present application. In Figure 8 In the embodiment shown, the signal change curve of the barometer g is shown, where the abscissa is time t, with the unit of s (seconds), and the ordinate is the signal value.

[0134] Suppose the second length of the second window is 4 seconds, the second step length is 1 second, and the second sampling frequency is 5 Hz, then:

[0135] In the first sampling period, since the second length is 4 s, the waveform of 0 s-4 s is intercepted in the waveform diagram, and since the second sampling frequency is 5 Hz, 5 sampling points are obtained in the 0 s-1 s period, the 1 s-2 s period, the 2 s-3 s period, and the 3 s-4 s period. The second sampling signal corresponding to the air pressure signal can be represented as T2*F2*N2, where T2 is the second length, F2 is the second sampling frequency, and N2 is the dimension of the air pressure signal, in this embodiment, N2 = 1. When the second length is 4 seconds, the second step length is 1 second, and the second sampling frequency is 5 Hz, the second sampling signal can be represented as a 4*5*1 vector.

[0136] In the second sampling period, since the second step length is 1 s, it is necessary to move 1 s, and the waveform of 1 s-5 s is intercepted in the waveform diagram, and then the second sampling signal of the second sampling period is obtained based on the same manner as the first sampling period. In this way, as time goes on, the counting period contained in the sliding window moves forward by 1 second, but always contains a 4-second period. Thus, the second sampling signal can be periodically obtained through the second window.

[0137] It should be understood that the embodiments of the present application do not limit the parameters of the second window, and the second length, the second step length and the second sampling frequency can be set according to the actual application scenario.

[0138] Step S320, obtaining a difference signal by performing difference processing on the second sampling signal through the first sub-window and the second sub-window.

[0139] The difference signal includes a first difference signal and a second difference signal, the first difference signal is obtained through the first sub-window, and the second difference signal is obtained through the second sub-window.

[0140] The first sub-window is a odd-odd window, the odd-odd window is a window for performing difference processing on a data point set with an odd time second number, and the second sub-window is an even-even window, the even-even window is a window for performing difference processing on a data point set with an even time second number.

[0141] Taking a second length of the second window as 4 seconds and a second sampling frequency as 5 Hz as an example, after a time window, the second sampling signal includes sampling values with a time length of 4 seconds, and the number of sampling points per second is 5. For the first sub-window, the values of the five sampling points of the third second are subtracted by the values of the five sampling points of the first second respectively, to obtain five difference values, and the first difference signal is obtained by summing the five difference values. Similarly, for the second sub-window, the values of the five sampling points of the fourth second are subtracted by the values of the five sampling points of the second second respectively, to obtain five difference values, and the second difference signal is obtained by summing the five difference values.

[0142] It should be understood that the parameters of the first sub-window and the second sub-window for difference processing in the embodiments of the present application are not limited, and can be set according to actual application scenarios.

[0143] Step S330, comparing the difference signal with a predetermined threshold to obtain the state category.

[0144] In the present embodiment, the state category includes at least one of an increase, a decrease, and a second category, wherein the second category is other than the increase and the decrease. The predetermined threshold includes a first threshold V th1 and a second threshold V th2 , wherein V th1 > 0 and V th2 < 0. In some embodiments, the first threshold V th1 and the second threshold V th2 are opposite numbers.

[0145] Specifically, comparing the first difference signal with the first threshold V th1 and the second difference signal with the second threshold V th2 includes:

[0146] Comparing the first difference signal and the second difference signal with the first threshold V th1 and the second threshold V th2 to obtain the state category.

[0147] Specifically, when the first differential signal is greater than the first threshold V th1 , and the second differential signal is greater than the first threshold V th1 , the state category is determined as falling.

[0148] When the first differential signal is less than the second threshold V th2 , and the second differential signal is less than the second threshold V th2 , the state category is determined as rising.

[0149] In other cases, the state category is determined as a second category. That is, when the first differential signal is greater than the first threshold V th1 , and the second differential signal is less than or equal to the first threshold V th1 , or when the first differential signal is less than or equal to the first threshold V th1 , and the second differential signal is greater than the first threshold V th1 , or when the first differential signal is less than the second threshold V th2 , and the second differential signal is greater than or equal to the second threshold V th2 , or when the second differential signal is greater than or equal to the second threshold V th2 , and the second differential signal is less than the first threshold V th1 , the state category is determined as the second category.

[0150] In step S400, a recognition result of the motion scene is determined according to the initial category and the state category.

[0151] In this embodiment, the initial category is at least one of walking and running, and the state category is at least one of rising and falling. The initial category and the state category are fused to determine the motion scene recognition result.

[0152] The motion scene recognition result includes:

[0153] When the initial category determines that the scene is walking, and the state category is a rising state, the motion scene recognition result is determined as walking down stairs.

[0154] When the initial category determines that the scene is running, and the state category is a rising state, the motion scene recognition result is determined as running down stairs.

[0155] When the initial category determines that the scene is walking, and the state category is a falling state, the motion scene recognition result is determined as walking up stairs.

[0156] When the initial category discriminates the scene as running and the state category is the descending state, the recognition result of the motion scene is determined as running downstairs.

[0157] When the initial category is the first category and the state category is the ascending state, the recognition result of the motion scene is determined as elevator downstairs.

[0158] When the initial category is the first category and the state category is the descending state, the recognition result of the motion scene is determined as elevator upstairs.

[0159] Further, when the initial category is cycling, driving or elliptical machine and the state category is the second category, the initial category is determined as the recognition result of the motion scene.

[0160] When the initial category is driving and the state category is the second category, the recognition result of the motion scene is determined as driving.

[0161] When the initial category is elliptical machine and the state category is the second category, the recognition result of the motion scene is determined as elliptical machine.

[0162] When the initial category is cycling and the state category is the second category, the recognition result of the motion scene is determined as cycling.

[0163] In some embodiments, the method further comprises:

[0164] Step S500, the motion scene recognition result is displayed by the display module.

[0165] In the embodiment, the motion scene recognition result displayed by the display module can be active display or passive display.

[0166] Active display means that the processor synchronizes the recognition result to the display module after obtaining the recognition result of the motion scene, so that the display module directly displays the recognition information of the motion scene.

[0167] Passive display means that when the processor obtains the recognition result of the motion scene, the recognition result is stored, and when the user triggers the display instruction (such as turning on the screen, clicking the display control of the motion scene, etc.), the recognition result of the motion scene is obtained and displayed in the stored data. Or, when the user triggers the display instruction (such as turning on the screen, clicking the display control of the motion scene, etc.), the processor obtains the recognition result of the motion scene, synchronizes the recognition result to the display module, so that the display module directly displays the recognition information of the motion scene.

[0168] The embodiment of the present application obtains a gravity acceleration signal and an air pressure signal, obtains an initial category of a motion scene according to the gravity acceleration signal, obtains a state category according to the air pressure signal, and determines a recognition result of the motion scene according to the initial category and the state category. Thus, through multi-sensor fusion processing, the operation amount and power consumption of motion scene recognition can be reduced, and the device endurance time and recognition accuracy can be improved.

[0169] Figure 9 is a schematic diagram of a motion scene recognition device according to an embodiment of the present application. As shown in Figure 9 the motion scene recognition device according to the embodiment of the present application includes a detection signal obtaining unit 91, an initial category obtaining unit 92, a state category obtaining unit 93, and a recognition result determining unit 94. The detection signal obtaining unit 91 is configured to obtain a detection signal, the detection signal including a gravity acceleration signal and an air pressure signal. The initial category obtaining unit 92 is configured to obtain an initial category of a motion scene according to the gravity acceleration signal. The state category obtaining unit 93 is configured to obtain a state category according to the air pressure signal. The recognition result determining unit 94 is configured to determine a recognition result of the motion scene according to the initial category and the state category.

[0170] In some embodiments, the detection signal obtaining unit includes:

[0171] an acceleration signal obtaining subunit configured to receive a gravity acceleration signal sent by a three-axis gravity acceleration sensor; and

[0172] an air pressure signal obtaining subunit configured to receive an air pressure signal sent by a barometer.

[0173] In some embodiments, the initial category obtaining unit includes:

[0174] a first sampling subunit configured to sample the gravity acceleration signal according to a predetermined first length, a first step length, and a first sampling frequency through the first window to obtain the first sampled signal; and

[0175] a classification subunit configured to obtain the initial category according to the first sampled signal through a pre-trained classification model.

[0176] In some embodiments, the pre-trained classification model includes an attention layer, a weighting layer, a first convolution layer, a second convolution layer, a merging layer, a third convolution layer, and a fully connected layer.

[0177] The classification subunit includes:

[0178] a first intermediate vector obtaining module configured to obtain a first intermediate vector according to the first sampled signal through the attention layer;

[0179] a second intermediate vector obtaining module configured to obtain a second intermediate vector from the first sampling signal and the first intermediate vector through the weighting layer;

[0180] a third intermediate vector obtaining module configured to obtain a third intermediate vector from the second intermediate vector through the first convolutional layer;

[0181] a fourth intermediate vector obtaining module configured to obtain a fourth intermediate vector from the third intermediate vector through the second convolutional layer;

[0182] a fifth intermediate vector obtaining module configured to obtain a fifth intermediate vector from the third intermediate vector and the fourth intermediate vector through the merging layer;

[0183] a sixth intermediate vector obtaining module configured to obtain a sixth intermediate vector from the fifth intermediate vector through the third convolutional layer; and

[0184] an initial category obtaining module configured to obtain the initial category from the sixth intermediate vector through the fully connected layer.

[0185] In some embodiments, the state category obtaining unit comprises:

[0186] a second sampling subunit configured to sample the air pressure signal according to a predetermined second length, a second step length, and a second sampling frequency through the second window to obtain the second sampling signal; and

[0187] a state obtaining subunit configured to obtain the state category from the second sampling signal.

[0188] In some embodiments, the obtaining state subunit comprises:

[0189] a differential signal obtaining module configured to obtain a differential signal by differentially processing the second sampling signal through a first sub-window and a second sub-window; and

[0190] a differential signal comparison module configured to compare the differential signal with a predetermined threshold to obtain the state category.

[0191] In some embodiments, the initial category comprises at least one of walking, running, cycling, driving, elliptical machine, and a first category, and the state category comprises at least one of going upstairs, going downstairs, and a second category.

[0192] In some embodiments, the motion scene recognition result determined by the recognition result determining unit comprises at least one of:

[0193] in response to the initial category being walking and the state category being going upstairs, determining that the recognition result of the motion scene is walking downstairs;

[0194] determining that the identification result of the motion scene is running down stairs in response to the initial category being running and the state category being ascending;

[0195] determining that the identification result of the motion scene is walking up stairs in response to the initial category being walking and the state category being descending;

[0196] determining that the identification result of the motion scene is running up stairs in response to the initial category being running and the state category being descending;

[0197] determining that the identification result of the motion scene is elevator down in response to the initial category being the first category and the state category being ascending;

[0198] determining that the identification result of the motion scene is elevator up in response to the initial category being the first category and the state category being descending;

[0199] determining that the initial category is the identification result of the motion scene in response to the initial category being cycling, driving or elliptical machine and the state category being the second category.

[0200] The embodiment of the application obtains a gravity acceleration signal and an air pressure signal, obtains an initial category of a motion scene according to the gravity acceleration signal, obtains a state category according to the air pressure signal, and determines an identification result of the motion scene according to the initial category and the state category. Thus, through multi-sensor fusion processing, the computational amount and power consumption of motion scene identification can be reduced, and the device endurance time and identification accuracy can be improved.

[0201] The CAM-TCN-net (Class Activation Mapping Temporal Convolutional Network) model for real-time motion scene identification is provided. The embodiment of the application is a collection device of three-axis gravity acceleration and barometer signals, which is verified by labeled sample data. The three-axis gravity acceleration signal corresponds to six category sample labels of walking, running, cycling, driving, elliptical machine and others, and the barometer signal corresponds to three category sample labels of ascending, descending and others. The accuracy performance indicators are Ac (Accuracy), Re (Recall), Sp (Specificity) and Mcc (Matthews correlation coefficient).

[0202] The calculation formula of the accuracy (Ac for short) is:

[0203]

[0204] wherein the calculation formula of recall (abbreviation Re) is:

[0205]

[0206] wherein the calculation formula of specificity (abbreviation Sp) is:

[0207]

[0208] wherein the calculation formula of Matthew correlation coefficient (abbreviation Mcc) is:

[0209]

[0210] wherein TP (True Positive, true positive), TN (True Negative, true negative), FP (False Positive, false positive) and FN (False Negative, false negative) are used for prediction value. The value range of Matthew correlation coefficient Mcc is from-1 to 1, Mcc is-1, indicating the worst possible prediction, and the value is 1, indicating the best possible prediction scheme. In addition, Mcc is 0, indicating random prediction. 10000 three-axis gravity acceleration sample points collected and labeled are divided into training set and test set according to the ratio of 8:2, wherein the sample points of the training set and the test set are 8000 and 2000, wherein the classification results of the test set of the CAB-TCN-net (Channel Attention Block-Convolutional Neural Network, network structure constructed based on attention mechanism neural network model) model after training are compared with the performance indicators of the single TCN (Temporal Convolutional Network, time sequence convolution network) and the similar structure CAB-CNN (Channel Attention Block-Convolutional Neural Network, convolution neural network model based on attention mechanism) and single CNN (Convolutional Neural Network, convolution neural network) as shown in Figure 10 It can be seen that compared with TCN, CAB-CNN and CNN model, the CAB-TCN-net proposed in the application has obvious performance improvement for the test set for real-time identification of motion scene.

[0211] 6000 barometer-based samples collected and labeled, including 1218 upstairs samples, 1321 downstairs samples and 3461 other samples, are directly identified using the upstairs and downstairs identification unit proposed in the application, and the identification classification results are represented by performance indicators asFigure 11 As shown in the table, each performance index reaches more than 99%, which indicates that the up-and-down floor recognition unit has good performance.

[0212] It should be noted that, Figure 1 In the electronic device shown, the processor 14 and the memory 13 are connected through a bus. The memory 13 is adapted to store instructions or programs executable by the processor 14. The processor 14 can be a stand-alone microprocessor or a set of one or more microprocessors. Thus, the processor 14 performs the processing of data and the control of other devices by executing the instructions stored in the memory 13, thereby implementing the method flow of the embodiments of the present application as described above. The bus connects the above components together, and connects the above components to a display controller and a display device and input / output (I / O) devices. The input / output (I / O) devices can be a mouse, a keyboard, a modem, a network interface, a touch input device, a body sense input device, a printer, and other devices known in the art. Typically, the input / output devices are connected to the system through an input / output (I / O) controller.

[0213] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, device (apparatus) or computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM (Compact Disc Read-Only Memory), optical storage, etc.) containing computer usable program code.

[0214] The present application is described with reference to flowcharts according to the methods, devices (apparatuses) and computer program products of the embodiments of the present application. It should be understood that each flow in the flowcharts can be implemented by computer program instructions.

[0215] These computer program instructions can be stored in a computer readable memory capable of directing the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices, which implement the functions specified in the flow Figure 1 or multiple flows.

[0216] These computer program instructions can also be provided to the processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in the flow Figure 1 or multiple flows.

[0217] The above descriptions are only the preferred embodiments of the present application, and are not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A moving scene recognition method, characterized in that, The method comprises: obtaining a detection signal, the detection signal comprising a gravity acceleration signal and an air pressure signal; obtaining an initial category of a motion scene according to the gravity acceleration signal; obtaining a state category according to the air pressure signal; and determining a recognition result of the motion scene according to the initial category and the state category; the obtaining of the state category according to the air pressure signal comprises: sampling the air pressure signal through a second window according to a predetermined second length, a second step and a second sampling frequency to obtain a second sampling signal; differentially processing the second sampling signal through a first sub-window and a second sub-window to obtain a differential signal; comparing the differential signal with a predetermined threshold to obtain the state category; wherein the first sub-window is a window for differentially processing a data point set with an odd time second, the second sub-window is a window for differentially processing a data point set with an even time second, the predetermined threshold comprises a first threshold greater than 0 and a second threshold less than 0, and the first threshold and the second threshold are opposite numbers.

2. The method of claim 1, wherein, the obtaining of the detection signal comprises: receiving a gravity acceleration signal sent by a three-axis gravity acceleration sensor; and receiving an air pressure signal sent by an air pressure gauge.

3. The method of claim 1, wherein, the obtaining of the initial category of the motion scene according to the gravity acceleration signal comprises: sampling the gravity acceleration signal through a first window according to a predetermined first length, a first step and a first sampling frequency to obtain a first sampling signal; and obtaining the initial category according to the first sampling signal through a pre-trained classification model.

4. The method of claim 3, wherein, the pre-trained classification model comprises an attention layer, a weighting layer, a first convolutional layer, a second convolutional layer, a merging layer, a third convolutional layer and a fully connected layer; wherein the obtaining of the initial category according to the first sampling signal through the pre-trained classification model comprises: obtaining a first intermediate vector according to the first sampling signal through the attention layer; obtaining a second intermediate vector according to the first sampling signal and the first intermediate vector through the weighting layer; generating a third intermediate vector according to the second intermediate vector through the first convolutional layer; generating a fourth intermediate vector according to the third intermediate vector through the second convolutional layer; generating a fifth intermediate vector according to the third intermediate vector and the fourth intermediate vector through the merging layer; generating a sixth intermediate vector according to the fifth intermediate vector through the third convolutional layer; and obtaining the initial category according to the sixth intermediate vector through the fully connected layer.

5. The method of claim 1, wherein, the initial category comprises at least one of walking, running, cycling, driving, elliptical machine and a first category, and the state category comprises at least one of going upstairs, going downstairs and a second category.

6. The method of claim 5, wherein, the determination of the recognition result of the motion scene according to the initial category and the state category comprises at least one of: in response to the initial category being walking and the state category being ascending, determining that the recognition result of the motion scene is walking downstairs; in response to the initial category being running and the state category being ascending, determining that the recognition result of the motion scene is running downstairs; In response to the initial category being walking and the state category being descending, determining that the identification result of the motion scene is walking upstairs; In response to the initial category being running and the state category being descending, determining that the identification result of the motion scene is running upstairs; In response to the initial category being the first category and the state category being ascending, determining that the identification result of the motion scene is elevator descending; In response to the initial category being the first category and the state category being descending, determining that the identification result of the motion scene is elevator ascending; In response to the initial category being cycling, driving or elliptical machine and the state category being the second category, determining that the initial category is the identification result of the motion scene.

7. A moving scene recognition apparatus characterized by comprising: The apparatus comprises: a detection signal acquisition unit configured to acquire a detection signal, the detection signal comprising a gravity acceleration signal and an air pressure signal; an initial category acquisition unit configured to acquire an initial category of a motion scene according to the gravity acceleration signal; a state category acquisition unit configured to acquire a state category according to the air pressure signal; and an identification result determination unit configured to determine an identification result of the motion scene according to the initial category and the state category; The state category acquisition unit is further configured to: sample the air pressure signal through a second window according to a predetermined second length, a second step and a second sampling frequency to acquire a second sampling signal; perform differential processing on the second sampling signal through a first sub-window and a second sub-window to acquire a differential signal; compare the differential signal with a predetermined threshold to acquire the state category; wherein the first sub-window is a window for performing differential processing on a data point set with an odd number of time seconds, the second sub-window is a window for performing differential processing on a data point set with an even number of time seconds, the predetermined threshold comprises a first threshold greater than 0 and a second threshold less than 0, the first threshold and the second threshold are opposite numbers.

8. An electronic device, comprising: The electronic device comprises: a three-axis gravity acceleration sensor configured to acquire a gravity acceleration signal; an air pressure gauge configured to acquire an air pressure signal; and a memory and a processor, the memory being configured to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1-6.

Images