A method for obtaining a human body perception model based on radio frequency technology and its utilization method

Abstract

A method for obtaining a human perception model based on radio frequency technology, belonging to the field of computer technology, includes steps S1: based on radio frequency scanning, point cloud data collection for a selected scene; S2: solving the point cloud information corresponding to each point cloud data group; S3: according to the reference information recorded during the data collection process, marking the spatial position information of the human representative points in the scene corresponding to each group of point cloud information obtained by radio frequency scanning, making a first sample set, and further using machine learning to train a model MO that can identify the number of people M in the scene and the spatial position of each human representative point. The method can also include the steps of making a second sample set, training a model MK, making a third sample set, and training a model MA. The method can be carried out without human perception, reducing the rejection of the human target being measured, and filtering the point cloud data with reference to the output result of the model MO, which can effectively reduce the amount of data calculation.

Description

Technical Field

[0001] The present application belongs to the field of computer technology, and specifically relates to a method for obtaining a human body perception model based on radio frequency technology and machine learning, and also relates to a method for performing human body perception using the corresponding model. Background Art

[0002] In traditional human body sensing solutions, the person being detected needs to actively provide data to the sensing system by wearing or donning various devices. If long-term monitoring is required, the person being detected needs to wear or don the corresponding equipment 24 hours a day, which increases the burden on the person being detected and affects their normal life.

[0003] In recent years, with the continuous promotion of the application of visual recognition technology, in terms of human perception, visual recognition technology can also perceive and judge information such as human movements. However, perceiving human movements or behaviors through visual recognition can easily affect the personal privacy of the person being detected and is difficult to promote on a large scale.

[0004] The point cloud information is obtained by scanning with laser radar and ultrasonic radar, and then the three-dimensional model of the scanned object is inverted, which has been widely used in the field of surveying and navigation. Point cloud is a massive set of points that express the spatial distribution and surface characteristics of the target in the same spatial reference system, and the amount of data is very large. Taking the device collecting 17.5M samples per second, 4 sets of sampling devices, and 16-bit complex sampling as an example, the amount of sampled data per second is as high as 1.6Gbps. At present, some similar algorithm designs and experimental bases are based on direct algorithm implementation using PCs and servers, which is extremely costly and seriously restricts the application of related technologies in the field of home monitoring. Summary of the invention

[0005] The present application aims to provide a method for acquiring a human body perception model based on radio frequency technology, and also provides a method for performing human body perception using a corresponding human body perception model, as well as corresponding electronic equipment.

[0006] To achieve the above objectives, this application adopts the following four technical solutions:

[0007] In a first aspect, the present application provides a method for obtaining a human perception model based on radio frequency technology, comprising the following steps:

[0008] S1: Based on RF scanning, point cloud data collection is performed on a selected scene;

[0009] S2: Consider a group of point cloud data obtained by N radio frequency scans as one point cloud data group, and solve the point cloud information corresponding to each point cloud data group, wherein the information corresponding to each reflection point in the point cloud information at least includes the spatial position, speed and signal strength information of the reflection point; N is an integer greater than or equal to 2;

[0010] S3: Based on the reference information recorded during the data collection process, mark the spatial position information of the representative points of the human body in the scene corresponding to each set of point cloud information obtained by the RF scan; collect multiple sets of point cloud information obtained by the RF scan and the corresponding spatial position information of the representative points of the human body to form a first sample set; based on the first sample set, use machine learning methods to train a model MO that can identify the number of people M in the scene and the spatial position of each representative point of the human body.

[0011] Furthermore, the output target of the model MO is O={(Pr m , Ps m ), m=1, 2, 3,..., M}, Pr m is the probability of the existence of the mth human target to be detected, Ps m is the spatial position of the representative point of the mth human target to be detected, and M is the number of people in the scene.

[0012] Furthermore, the method further comprises step S4:

[0013] N pk Group (N pk The point cloud information obtained by the radio frequency scanning (N times) is regarded as a point cloud information sequence, and the point cloud information sequence is filtered according to the output result of the model MO, and only the point cloud information within a specific distance range near the representative point of the human body is retained to obtain the filtered point cloud information sequence; pk is an integer greater than 1;

[0014] Select multiple key points on the human body based on the human joints, mark the spatial position information of the human key points corresponding to each filtered point cloud information sequence according to the reference information recorded in the formation of the data set, collect multiple information sequences and filter and mark them to form a second sample set;

[0015] Based on the second sample set, a model MK is trained using a machine learning method, which can recognize multiple key point information of M human bodies in the scene.

[0016] Furthermore, the output target of the model MK is the key point information OK of each human target to be detected = {(Pr k , Ps k ), k = 1, 2, 3, ..., K}, K is the number of selected human key points; Pr k Ps is the probability of the kth key point of a human target to be detected; k It is the spatial position of the kth key point of a human target to be inspected.

[0017] Furthermore, the method further comprises step S5:

[0018] Obtain a third sample set, the third sample set includes positive samples and negative samples, the positive samples include a certain human behavior obtained from the reference information and the corresponding N ma The continuous output results of MK are used as counterexample samples where a certain behavior does not occur. ma is an integer greater than 1;

[0019] Based on the third sample set, a model MA is trained using machine learning methods, which can recognize specific human behaviors based on multiple point cloud information sequences.

[0020] Furthermore, the output target of the model MA is the probability of a certain human target to be detected to have a certain specific behavior.

[0021] Furthermore, the RF signal frequency is 3Ghz-90Ghz, and the bandwidth is 500Mhz-20Ghz.

[0022] Furthermore, before performing step S3, the point cloud data obtained by the radio frequency scanning is filtered according to the data obtained by the radio frequency scanning of the unmanned scene to filter out the fixed scene information.

[0023] In a second aspect, a method for human body perception using a model obtained by any implementation of the first aspect comprises the following steps:

[0024] S10: Based on radio frequency scanning, point cloud data collection of the scene is performed;

[0025] S20: Considering a group of point cloud data obtained by N radio frequency scans as one point cloud data group, and solving point cloud information corresponding to each point cloud data group, wherein the information corresponding to each reflection point P in the point cloud information at least includes the spatial position, speed and signal strength information of the reflection point;

[0026] S30: Input the point cloud information calculated in step S20 into the model MO, and output the number of people in the scene and the spatial position information of each representative point of the human body.

[0027] Furthermore, the method for human body perception using a model further includes step S40: when the number of human bodies M output by MO is ≥ 1, the point cloud information obtained by radio frequency scanning is filtered according to the spatial position information of the human body representative points output by the model MO, and only the point cloud information within a specific distance range near the human body representative points is retained; the filtered point cloud information is input into the model MK, and the model MK uses a sliding window method to scan and identify the input point cloud information, and the window length corresponds to N pk A group of radio frequency scans are performed to output multiple key point information of M human bodies.

[0028] Furthermore, the method for human body perception using a model also includes the step S50: inputting the output result of the model MK into the model MA, and the model MA uses a sliding window method to scan and identify the input information, and the window length corresponds to N ma The MK continuously outputs the results and outputs specific human behaviors.

[0029] Furthermore, a radio frequency scanning cycle T c =20~2000μs, and each point cloud data group corresponds to the number of RF scans N=2~4096.

[0030] The reference information is a video record or an audio and video record obtained synchronously during the radio frequency scanning process; the training set can be marked manually. As a more preferred solution, the existing artificial intelligence recognition method can be used to extract the representative point positions, key point positions and human behavior information of the human body from the reference information, and then automatically mark the point cloud information based on the same time axis.

[0031] To improve the efficiency and accuracy of machine learning, the point cloud information corresponding to each reflection point may further include acceleration and noise amplitude information.

[0032] In a third aspect, the present application provides an electronic device, including:

[0033] one or more processors;

[0034] a storage device having one or more programs stored thereon;

[0035] When the one or more programs are executed by the one or more processors, the one or more processors implement any method described in the first aspect or any method described in the second aspect.

[0036] In a fourth aspect, the present application provides a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the program implements any of the methods described in the first aspect or any of the methods described in the second aspect.

[0037] Compared with the prior art, the present application can sense the position, movement or behavior of the person being tested without the person being aware of it, unlike monitoring methods such as video surveillance that are prone to infringe on personal privacy. Specifically, the radio frequency signal is used to form a radio frequency wave field, and then point cloud information that expresses the spatial distribution of the scene to be tested and the surface characteristics of the human target to be tested is obtained; the obtained point cloud information and reference information are used to produce a first sample set, and a model MO that can identify the position of representative points of the human body is trained through machine learning; further, the output result of MO is used to filter the obtained point cloud information, and only highly correlated information is retained to produce a second sample set, which greatly reduces the computing burden of the device. Based on the second sample set, a model MK that can identify the key point information of the human body is trained, and a third sample set is produced based on the output result of MK, and the model MA is further trained, which further reduces the computing requirements and can effectively reduce the equipment cost. Accordingly, the trained model is used for human perception, and the point cloud information is first filtered using MO and / or MK, which helps to distinguish between the human body and other moving objects such as fans, pets, etc., and improve the accuracy of the next level of recognition; further, the application of MO, MK, and MA step-by-step reasoning can greatly save the amount of computing power, reduce the burden on users, and promote the widespread application of radio frequency sensing or recognition technology in public nursing homes, hospitals, and especially families or home-based nursing facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of Example 1;

[0039] Figure 2 It is a layout diagram of the key points of the human body;

[0040] Figure 3 is a flow chart of Example 2;

[0041] Figure 4 This is the schematic diagram of the sliding window. DETAILED DESCRIPTION

[0042] The present application is further described below in conjunction with the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] A method for obtaining a human body perception model based on radio frequency technology, Figure 1 A process of the method is shown, which may include the following steps:

[0045] S1: Based on RF scanning, point cloud data is collected for a selected scene. The acceptable frequency range of RF signals is 3Ghz-90Ghz, and the bandwidth is 500Mhz-20Ghz. The transmission and reception of RF signals can be achieved through MIMO antennas pre-installed in the scene. In order to obtain stereo signals, multiple groups of antennas can be arranged in the scene to obtain gridded point cloud data. During the RF scanning process, the scene is recorded synchronously or other marking methods are used to obtain reference information.

[0046] S2: A set of point cloud data obtained by N times (N is an integer greater than or equal to 2) of radio frequency scanning is regarded as a point cloud data group, and the point cloud information corresponding to each point cloud data group is solved; the information corresponding to each reflection point P in the point cloud information includes at least the spatial position (x, y, z), velocity v (velocity information can be obtained when N is greater than or equal to 2) and signal strength g information of the reflection point, and can further include acceleration a (acceleration information can be obtained when N is greater than or equal to 3) and noise amplitude n, denoted as P{(x, y, z), v, g, a, n}.

[0047] To obtain the spatial position information of the target to be detected, the device needs to linearly scan the B bandwidth frequency band within the Tc time period, transmit the radio frequency and receive the radio frequency at the same time, mix the two and filter out the high-frequency signal to obtain the intermediate frequency signal, and then sample the signal. Since the scanning frequency increases linearly, , where τ is the time required for the transmitted signal to travel back and forth from the device to the target to be detected, Tc is the time period, ƒτ is the frequency of the received intermediate frequency signal, and B is the bandwidth band. The distance between the target to be detected and the device is d= , where C is the speed of light. By performing Fourier transform on the sampled signal, the ƒτ value of the reflection point is obtained, and then the distance information of the reflection point, i.e. the target to be detected, is obtained.

[0048] The acceleration information of the target to be detected is obtained. At this time, the target to be detected is in a moving state. Due to the existence of the Doppler phenomenon, the phase of the radio frequency received during the two detections will change greatly. The displacement between the two reflection points of the target to be detected can be obtained through the phase change: , the instantaneous speed is ,in is the wavelength of the radio frequency used, The phase difference between the two scans is obtained by scanning at least three times to obtain the acceleration of the target at each reflection point. c =20~3500μs.

[0049] In order to reduce the burden of subsequent data calculations, during the solution process, the point cloud data obtained by radio frequency scanning can be filtered according to the data obtained by radio frequency scanning of unmanned scenes to filter out fixed scene information.

[0050] S3: based on the reference information recorded during the data collection process, mark the spatial position information of the human body representative point corresponding to each set of point cloud information obtained by the radio frequency scanning in the scene. As an implementation method, the center point of the human body trunk can be selected as the representative point; when marking, the position of the human body representative point can be extracted from the reference information using the existing artificial intelligence recognition method, and then the point cloud information can be automatically marked based on the same time axis;

[0051] Collecting multiple sets of point cloud information obtained by radio frequency scanning and corresponding spatial position information of human body representative points to form a first sample set;

[0052] Based on the first sample set, a model MO capable of identifying the number of people M and the spatial position of each representative point of the human body in the scene is trained using machine learning methods such as random forest, support vector machine, AdaBoost or Gradient Tree Boosting based on decision tree, neural network, etc. The output target of the model MO is O={(Pr m , Ps m ), m=1, 2, 3,..., M}, Pr m is the probability of the existence of the mth human target to be detected, Ps m is the spatial position of the representative point of the mth human target to be detected, and M is the number of people in the scene. According to the selected algorithm, numerical loss functions such as MSE and McHatton distance between input and output values ​​are used as evaluation methods to improve the accuracy of the model.

[0053] S4: N pk Group (N pk ×N times, N pk is an integer greater than 1, preferably 2 to 25) the point cloud information obtained by radio frequency scanning is regarded as a point cloud information sequence, and the point cloud information sequence is filtered according to the output result of the model MO, and only the point cloud information within a specific distance range or even a specific speed range near the representative point of the human body is retained (the data within the size range of the human body is taken as valid data, which is conducive to further reducing the amount of data calculation) to obtain a filtered point cloud information sequence;

[0054] Based on the human joints, multiple key points are selected on the human body. According to the reference information recorded in the data set formation, the spatial position information of the human key points corresponding to each filtered point cloud information sequence is marked. Multiple information sequences are collected and filtered and marked to form a second sample set. The selection of human key points can refer to Figure 2 ,exist Figure 2 In the embodiment shown, K=8, Figure 2The serial numbers 1-8 in the figure indicate eight key points of the human body, namely, torso 1 (same as the representative point of the human body), head 2, elbows 3 and 4, knees 5 and 6, hands 7 and 8;

[0055] Based on the second sample set, a model MK capable of identifying multiple key point information of M human bodies in the scene is trained by using machine learning methods, such as random forest, support vector machine, AdaBoost or Gradient Tree Boosting based on decision tree, neural network, etc. The output target of the model MK is the key point information OK of each human body target to be detected = {(Pr k , Ps k ), k = 1, 2, 3, ..., K}, K is the number of selected human key points; Pr k Ps is the probability of the kth key point of a human target to be detected; k is the spatial position of the kth key point of a human target to be detected. According to the selected algorithm, numerical loss functions such as MSE and McHatton distance between input and output values ​​are used as evaluation methods to improve the accuracy of the model.

[0056] S5: Obtain a third sample set, the third sample set includes positive samples and negative samples, the positive samples include a certain human behavior (for example, falling) obtained from the reference information and the corresponding N ma times (N ma is an integer greater than 1, preferably 18 to 750) continuous MK output results, and the remaining MK output results are used as counterexample samples where a certain specific behavior does not occur;

[0057] Based on the third sample set, a model MA capable of identifying specific human behaviors based on multiple point cloud information sequences is trained using machine learning methods, such as random forest, support vector machine, AdaBoost or Gradient Tree Boosting based on decision tree, neural network, etc.; the output target of the model MA is the probability of a certain human target to be detected having a certain specific behavior. According to the selected algorithm, a category loss function such as cross entropy in neural network or Hinge in support vector machine is used as an evaluation method to improve the model accuracy.

[0058] Similarly, refer to step S5 to set other specific human behaviors, such as running, jumping, etc., and adjust the parameter N according to the duration of the action. ma , repeating step S5, multiple models MA that can recognize different human behaviors can be obtained.

[0059] As an implementation method, T c =1000μs, N=3, N pk =10,Nma =50, monitor the falling behavior through MA, monitor N ma The corresponding time of this data is 1500ms, which is basically the same as the time required for a fall.

[0060] In order to improve the wide application of the model, different settings can be made for the scenarios, and different numbers of people can be arranged to perform different activities in the scenarios, so as to obtain a richer sample set.

[0061] Example 2

[0062] A method for human body perception using the model obtained in Example 1, Figure 3 A process of the method is shown, which may include the following steps:

[0063] S10: Based on radio frequency scanning, point cloud data collection of the scene is performed;

[0064] S20: Considering a set of point cloud data obtained by multiple radio frequency scanning as one point cloud data group, and solving point cloud information corresponding to each point cloud data group, wherein the information corresponding to each reflection point P in the point cloud information at least includes the spatial position, speed and signal strength information of the reflection point;

[0065] S30: input the point cloud information calculated in step S2 into the model MO, and output the number of people in the scene and the spatial position information of each human representative point;

[0066] S40: Determine the output result of MO. When the number of human bodies M output by MO is ≥ 1, filter the point cloud information obtained by radio frequency scanning according to the spatial position information of the representative points of the human body output by the model MO, and only retain the point cloud information within a specific distance range or even a specific speed range near the representative points of the human body. Input the filtered point cloud information into the model MK and start the model MK. The model MK uses the sliding window method to scan and identify the input point cloud information. The window length corresponds to N. pk A group of radio frequency scans are performed to output multiple key point information of M human bodies. The principle of the sliding window is as follows Figure 4 As shown, in Figure 4 In the embodiment shown, the window length corresponds to N pk =10, pane width S pk=2, that is, when the model MK receives 4 sets of point cloud information obtained by radio frequency scanning, it forms the initial state of the window, corresponding to step S401, and performs a scan and recognition; MK continues to receive 2 sets of point cloud information obtained by radio frequency scanning, the window slides forward once, subtracts the first 2 sets of invalid information and adds 2 sets of latest information to form the current window, corresponding to step S402, for the current window information, MK performs a second scan and recognition; MK continues to receive 2 sets of point cloud information obtained by radio frequency scanning, the window slides forward once, subtracts the first 2 sets of invalid information in the S402 window and adds 2 sets of latest information to form a new current window again, corresponding to step S403, performs a third scan and recognition for the information corresponding to the new window; and so on, traverse all received information. Of course, S pk It can take a smaller value, such as S pk =1, or a greater integer, S pk The larger the value, the lower the computing burden of the device, but the recognition accuracy will also decrease. When the human body is in a state of low activity such as sleep, S can be appropriately increased. pk The value of, for example, S pk =N pk / 2=5 or S pk =N pk =10.

[0067] S50: Input the output result of the model MK to one or more models MA. The model MA uses a sliding window method to scan and identify the input information. The window length corresponds to N ma The MK continuously outputs the results and outputs specific human behavior. The working principle of the sliding window is consistent with that described in S40, but as an implementation method, the window width S in step S50 is ma Preferably N ma / 2 (same as Example 1, N ma =50). It is understandable that S ma It can also be appropriately reduced, in exchange for accuracy through calculation, or increased, possibly sacrificing accuracy in exchange for speed. The models MA used to identify multiple different human behaviors run synchronously, and each model MA scans and identifies at its own set interval, infers and identifies the possibility of different behaviors, and outputs them.

[0068] Furthermore, a step of judging the human behavior type output by the model MA can be added, and the window width S can be set according to different human behaviors. pk The specified value is fed back to step S40, and the window width S is given according to different human behaviors. ma The specified value is fed back to step S50.

[0069] The above-mentioned human perception method is used to perform human perception in scenes such as homes, nursing homes, and hospitals. The model can continuously output the number of people in the scene, their locations, and key points of the human body, and identify specific behaviors of people. When the model recognizes that someone has fallen or performed dangerous actions, it can further transmit signals to the alarm system to remind guardians to take timely measures.

[0070] Example 3

[0071] An electronic device, comprising:

[0072] one or more processors;

[0073] a storage device having one or more programs stored thereon;

[0074] When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in Example 1.

[0075] Example 4

[0076] An electronic device, comprising:

[0077] one or more processors;

[0078] a storage device having one or more programs stored thereon;

[0079] When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in Example 2.

[0080] Example 5

[0081] A computer-readable medium stores a computer program, wherein the program implements the method described in embodiment 1 when executed by a processor.

[0082] Example 6

[0083] A computer-readable medium having a computer program stored thereon, wherein the program implements the method described in Example 2 when executed by a processor.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for obtaining a human perception model based on radio frequency technology, characterized in that: The following steps are involved: S1: Based on RF scanning, point cloud data collection is performed on a selected scene; S2: Consider a group of point cloud data obtained by N radio frequency scans as one point cloud data group, and solve the point cloud information corresponding to each point cloud data group, wherein the information corresponding to each reflection point in the point cloud information at least includes the spatial position, speed and signal strength information of the reflection point; N is an integer greater than or equal to 2; S3: based on the reference information recorded during the data collection process, mark the spatial position information of the human body representative points in the scene corresponding to each set of point cloud information obtained by the radio frequency scanning; collect multiple sets of point cloud information obtained by the radio frequency scanning and the corresponding spatial position information of the human body representative points to form a first sample set; based on the first sample set, use the machine learning method to train a model MO that can identify the number of people M in the scene and the spatial position of each human body representative point; S4: N pk The point cloud information obtained by the group RF scanning is regarded as a point cloud information sequence, and the point cloud information sequence is filtered according to the output result of the model MO, and only the point cloud information within a specific distance range near the representative point of the human body is retained to obtain the filtered point cloud information sequence; N pk is an integer greater than 1; Select multiple key points on the human body based on the human joints, mark the spatial position information of the human key points corresponding to each filtered point cloud information sequence according to the reference information recorded in the formation of the data set, collect multiple information sequences and filter and mark them to form a second sample set; Based on the second sample set, a model MK is trained using a machine learning method, which can recognize multiple key point information of M human bodies in the scene.

2. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 1, characterized in that: The output target of the model MO is O={(Pr m , Ps m ), m=1, 2, 3,..., M}, Pr m is the probability of the existence of the mth human target to be detected, Ps m is the spatial position of the representative point of the mth human target to be detected, and M is the number of people in the scene.

3. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 1 or 2, characterized in that: The output target of the model MK is the key point information OK of each human target to be detected. k , Ps k ), k = 1, 2, 3, ..., K}, K is the number of selected human body key points; Pr k Ps is the probability of the kth key point of a human target to be detected; k It is the spatial position of the kth key point of a human target to be inspected.

4. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 1 or 2, characterized in that: The method further comprises step S5: Obtain a third sample set, the third sample set includes positive samples and negative samples, the positive samples include a certain human behavior obtained from the reference information and the corresponding N ma The continuous output results of MK are used as counterexample samples where a certain behavior does not occur. ma is an integer greater than 1; Based on the third sample set, a model MA is trained using machine learning methods, which can recognize specific human behaviors based on multiple point cloud information sequences.

5. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 4, characterized in that: The output target of the model MA is the probability of a certain human target to be detected to perform a certain specific behavior.

6. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 5, characterized in that: The frequency of the radio frequency signal is 3Ghz-90Ghz, and the bandwidth is 500Mhz-20Ghz.

7. The method for obtaining a human perception model based on radio frequency technology as claimed in claim 6, characterized in that: Before performing step S3, the point cloud data obtained by radio frequency scanning is first filtered according to the point cloud data obtained by radio frequency scanning of unmanned scenes to filter out fixed scene information.

8. A method for human body perception using the model obtained in claim 1 or 2 or 5 or 6 or 7, characterized in that: The following steps are involved: S10: Based on radio frequency scanning, point cloud data collection of the scene is performed; S20: Considering a group of point cloud data obtained by N radio frequency scans as one point cloud data group, and calculating point cloud information corresponding to each point cloud data group, wherein the information corresponding to each reflection point in the point cloud information at least includes the spatial position, speed and signal strength information of the reflection point; S30: Input the point cloud information calculated in step S20 into the model MO, and output the number of people in the scene and the spatial position information of each representative point of the human body.

9. The method for human body perception as claimed in claim 8, characterized in that: The step S40 is also included: when the number of human bodies M output by the MO is greater than or equal to 1, the point cloud information obtained by the radio frequency scanning is filtered according to the spatial position information of the representative points of the human body output by the model MO, and only the point cloud information within a specific distance range near the representative points of the human body is retained; the filtered point cloud information is input into the model MK, and the model MK uses a sliding window method to scan and identify the input point cloud information, and the window length corresponds to N pk A group of radio frequency scans are performed to output multiple key point information of M human bodies.

10. The method for human body perception according to claim 9, characterized in that: The step S50 is also included: the output result of the model MK is input into the model MA, and the model MA uses a sliding window method to scan and identify the input information. The window length corresponds to N ma The MK continuously outputs the results and outputs specific human behaviors.

11. An electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in claim 1 or 2 or 3 or 4 or 6 or 7, or implement the method as claimed in any one of claims 8-10.

12. A computer readable medium having a computer program stored thereon, wherein: When the program is executed by a processor, the method according to claim 1 or 2 or 3 or 4 or 6 or 7 is implemented, or the method according to any one of claims 8-10 is implemented.

Images