Curvature recognition method based on multi-angle LSTM neural network

Abstract

The application relates to a curvature recognition method based on a multi-angle LSTM neural network, which utilizes the periodic change of the distribution of intermediate frequency signals collected by a radar in an IQ domain when the incidence angle of radar signals changes to complete recognition of the curvature of an object surface; the method combines a radar positioning technology and utilizes an SVMD algorithm to complete filtering operation, so that the recognition of the curvature of the surface of a target object can be realized even in a complex environment; compared with the prior art, the method of the application does not have the problem of privacy leakage and can normally work in rain, fog and dark light environments; in addition, the method of the application can realize 5mm granularity object curvature recognition, and the object curvature recognition precision is relatively high.

Description

Technical Field

[0001] This application relates to the field of radio frequency sensing, and more specifically, to a curvature recognition method based on a multi-angle LSTM neural network. Background Technology

[0002] Wireless object recognition technology has evolved to the point where it offers new ways for people to perceive life and has become an important research direction, playing a crucial role in many related applications, such as autonomous driving, intelligent logistics, and assisted navigation for the blind or humanoid robots. However, these real-world applications often require the rapid and efficient differentiation of highly similar complex details. To address these challenges, there is a desire to find precise structural features to better achieve object recognition. The surface curvature of an object contains complex surface information and can help improve object recognition.

[0003] Existing methods for object surface curvature recognition include visual methods, but these methods suffer from privacy concerns and their performance deteriorates in environments such as rain, fog, and low light. RSA and Ulysses both utilize wireless signals for object curvature recognition, which addresses the privacy issues and performance degradation problems associated with visual systems in rain and fog. However, their curvature recognition error remains around 5cm, failing to achieve fine-grained object curvature recognition. Summary of the Invention

[0004] To overcome at least one deficiency in the prior art, this application provides a curvature recognition method based on a multi-angle LSTM neural network.

[0005] Firstly, a curvature recognition method based on a multi-angle LSTM neural network is provided, including:

[0006] The radar transmits signals at multiple incident angles toward the target object, and the radar receives the signals reflected by the target object. Based on the signals reflected by the target object, the radar obtains the intermediate frequency signals at multiple incident angles.

[0007] The intermediate frequency signals under multiple incident angles are filtered to obtain the filtered intermediate frequency signals under multiple incident angles;

[0008] Calculate the point density characteristics of the intermediate frequency signal under multiple incident angles based on the filtered intermediate frequency signal under multiple incident angles;

[0009] The point density features of intermediate frequency signals under multiple incident angles are sorted in ascending order of incident angle to obtain the sequence of changes in the point density features of the target object with the incident angle of the radar transmitted signal.

[0010] The changing sequence is input into a multi-angle LSTM neural network model, which outputs the radius of curvature of the target object.

[0011] In one embodiment, filtering is performed on intermediate frequency signals at multiple incident angles to obtain filtered intermediate frequency signals at multiple incident angles, including:

[0012] Determine the frequency value of the frequency component corresponding to the target object in the spectrum of the multipath environment;

[0013] The intermediate frequency (IF) signal is subjected to SVMD filtering, which decomposes the IF signal into multiple IMF signals;

[0014] The energy levels of each frequency component contained in each IMF signal are sorted to determine the frequency range in which the main energy of the IMF signal is located.

[0015] Select IMF signals that include the frequency value corresponding to the target object within the frequency range of the main energy source as the filtered IMF signals;

[0016] The filtered intermediate frequency signal is obtained by reconstructing the selected IMF signal.

[0017] In one embodiment, determining the frequency value of the frequency component corresponding to the target object in the spectrum of a multipath environment includes:

[0018] The distance-energy spectrum is obtained by applying the FFT algorithm to the intermediate frequency signal;

[0019] The CFAR algorithm is used to extract the distance information d between the target object and the radar antenna from the range-energy spectrum. ra ;

[0020] The frequency value is calculated using the following formula: Where K is the slope of the frequency-modulated continuous wave and c is the speed of light.

[0021] In one embodiment, the point density characteristics of the intermediate frequency signal at multiple incident angles are calculated based on the filtered intermediate frequency signal at multiple incident angles using the following formula:

[0022]

[0023] Where F is the point density feature of the intermediate frequency signal at each incident angle, N is the number of sampling points of the filtered intermediate frequency signal, and S is the area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain.

[0024] In one embodiment, a multi-angle LSTM neural network model includes:

[0025] The system consists of a first LSTM layer, a BatchNormalization layer, a second LSTM layer, and a fully connected layer.

[0026] Secondly, a curvature recognition device based on a multi-angle LSTM neural network is provided, comprising:

[0027] The intermediate frequency signal acquisition module is used for the radar to transmit signals at multiple incident angles to the target object, the radar to receive the signals reflected by the target object, and to acquire intermediate frequency signals at multiple incident angles based on the signals reflected by the target object.

[0028] The filtering module is used to filter intermediate frequency signals under multiple incident angles to obtain filtered intermediate frequency signals under multiple incident angles.

[0029] The point density feature calculation module is used to calculate the point density features of the intermediate frequency signal at multiple incident angles based on the filtered intermediate frequency signal at multiple incident angles.

[0030] The point density feature change sequence acquisition module is used to sort the point density features of intermediate frequency signals under multiple incident angles according to the incident angle from small to large, and obtain the change sequence of the point density features of the target object with the incident angle of the radar transmitted signal.

[0031] The radius of curvature acquisition module is used to input the changing sequence into the multi-angle LSTM neural network model and output the radius of curvature of the target object.

[0032] In one embodiment, the filtering module is further configured to:

[0033] Determine the frequency value of the frequency component corresponding to the target object in the spectrum of the multipath environment;

[0034] The intermediate frequency (IF) signal is subjected to SVMD filtering, which decomposes the IF signal into multiple IMF signals;

[0035] The energy levels of each frequency component contained in each IMF signal are sorted to determine the frequency range in which the main energy of the IMF signal is located.

[0036] Select IMF signals that include the frequency value corresponding to the target object within the frequency range of the main energy source as the filtered IMF signals;

[0037] The filtered intermediate frequency signal is obtained by reconstructing the selected IMF signal.

[0038] In one embodiment, the filtering module is further configured to:

[0039] The distance-energy spectrum is obtained by applying the FFT algorithm to the intermediate frequency signal;

[0040] The CFAR algorithm is used to extract the distance information d between the target object and the radar antenna from the range-energy spectrum. ra ;

[0041] The frequency value is calculated using the following formula: Where K is the slope of the frequency-modulated continuous wave and c is the speed of light.

[0042] In one embodiment, the point density feature calculation module is further configured to calculate the point density features of the intermediate frequency signal at each incident angle using the following formula:

[0043]

[0044] Where F is the point density feature of the intermediate frequency signal at each incident angle, N is the number of sampling points of the filtered intermediate frequency signal, and S is the area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain.

[0045] In one embodiment, a multi-angle LSTM neural network model includes:

[0046] The system consists of a first LSTM layer, a BatchNormalization layer, a second LSTM layer, and a fully connected layer.

[0047] Compared with existing technologies, this application has the following advantages: This application utilizes the periodic change in the distribution of the intermediate frequency signal collected by the radar in the IQ domain when the incident angle of the radar signal changes to complete the identification of the surface curvature of the object; the method combines radar positioning technology and uses the SVMD algorithm to complete the filtering operation, so that the surface curvature of the target object can be identified even in complex environments; compared with existing technologies, the method of this application does not have privacy leakage issues and can work normally in environments such as rain, fog and low light; in addition, the method of this application can achieve object curvature identification with a granularity of 5mm, and the object curvature identification accuracy is high. Attached Figure Description

[0048] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings:

[0049] Figure 1 A flowchart of a curvature recognition method based on a multi-angle LSTM neural network according to an embodiment of this application is shown;

[0050] Figure 2 The diagram illustrates a scenario where a millimeter-wave radar transmits and receives signals to a target object.

[0051] Figure 3 The result of filtering the intermediate frequency signal is shown in the figure.

[0052] Figure 4 This diagram illustrates how the distribution of the intermediate frequency signal reflected from an object changes in the IQ domain as the incident angle of the radar signal changes.

[0053] Figure 5 Experimental results are shown for the IQ domain distribution of intermediate frequency signals reflected from objects with different surface curvatures as a function of the incident angle of radar waves; where (a) is the experimental result for a concave object, (b) is the experimental result for a planar object, (c) is the experimental result for a convex object, and (d) is the experimental result for a display.

[0054] Figure 6 A schematic diagram illustrating the calculation of the area occupied by the signal cluster in the IQ domain is shown;

[0055] Figure 7 The experimental results show the point density characteristics of the intermediate frequency signals reflected from objects with different surface curvatures as a function of the radar incident angle.

[0056] Figure 8 A structural block diagram of a curvature recognition device based on a multi-angle LSTM neural network according to an embodiment of this application is shown;

[0057] Figure 9 The experimental results of curvature recognition using the method of this application under different environments are shown in the figure;

[0058] Figure 10 The experimental results of curvature recognition using the method of this application at different distances are shown in the figure;

[0059] Figure 11 The experimental results of curvature recognition using the method of this application at different angles are shown in the figure.

[0060] Figure 12 The experimental results of curvature recognition using the method of this application under different radar motion states are shown in the figure. Detailed Implementation

[0061] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0062] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution according to this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0063] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0064] This application addresses the privacy leakage issues, reduced recognition performance in rainy, foggy, or low-light environments, and insufficient accuracy of existing curvature recognition technologies. It proposes a curvature recognition method based on a multi-angle long short-term memory (LSTM) neural network. A millimeter-wave radar transmits and receives signals at both incident angles relative to the object. The positioning function of the millimeter-wave radar is then used to assist the Successive Variational Mode Decomposition (SVMD) algorithm in filtering under complex environments. Point density features are extracted at each angle, and a point density feature sequence is formed based on the angle magnitude. This sequence is then combined with a multi-angle LSTM neural network to achieve fine-grained object surface curvature recognition.

[0065] Figure 1 A flowchart of a curvature recognition method based on a multi-angle LSTM neural network according to an embodiment of this application is shown. The method includes:

[0066] Step S1: The radar transmits signals at multiple incident angles to the target object, and the radar receives the signals reflected by the target object. Based on the signals reflected by the target object, the radar obtains the intermediate frequency signals at multiple incident angles.

[0067] In this step, the transmission and reception of signals can be completed based on the millimeter-wave radar at two incident angles relative to the target object. Figure 2 The diagram illustrates a scenario where a millimeter-wave radar transmits and receives signals towards a target object. The reflected signals collected by the millimeter-wave radar are processed by the radar's internal mixer hardware to obtain an intermediate frequency (IF) signal.

[0068] Step S2: Filter the intermediate frequency signals under multiple incident angles to obtain filtered intermediate frequency signals under multiple incident angles;

[0069] Step S3: Calculate the point density characteristics of the intermediate frequency signals at multiple incident angles based on the filtered intermediate frequency signals at multiple incident angles.

[0070] Step S4: Sort the point density features of the intermediate frequency signals under multiple incident angles in ascending order of incident angle to obtain the sequence of changes in the point density features of the target object with the incident angle of the radar transmitted signal.

[0071] Step S5: Input the changing sequence into the multi-angle LSTM neural network model and output the radius of curvature of the target object.

[0072] In one embodiment, Figure 3 The diagram shows the result of filtering the intermediate frequency (IF) signal. In step S2, the IF signal under multiple incident angles is filtered to obtain filtered IF signals under multiple incident angles, including:

[0073] Step S21: Determine the frequency value of the frequency component corresponding to the target object in the spectrum of the multipath environment;

[0074] In this step, the FFT algorithm can be applied to the intermediate frequency signal first to obtain the distance-energy spectrum;

[0075] Then, the CFAR algorithm is used on the range-energy spectrum to extract the distance information d between the target object and the radar antenna. ra ;

[0076] The frequency value is calculated using the following formula: Where K is the slope of the frequency-modulated continuous wave and c is the speed of light.

[0077] Step S22: Perform SVMD filtering on the intermediate frequency signal to decompose the intermediate frequency signal into multiple IMF (Intrinsic Mode Function) signals;

[0078] Step S23: Sort the energy of each frequency component contained in each IMF signal according to its magnitude, and determine the frequency range in which the main energy of the IMF signal is located.

[0079] Step S24: Select IMF signals that include the frequency value corresponding to the target object within the frequency range of the main energy as the filtered IMF signals;

[0080] Step S25: Reconstruct the filtered intermediate frequency signal based on the selected IMF signal.

[0081] Furthermore, considering that the distribution of the intermediate frequency signal reflected from the object in the IQ domain (In-phase and Quadrature domain) changes periodically with the change of the radar signal incident angle, the principle behind this phenomenon is as follows:

[0082] Assuming an object is deployed in a static scene, its influence on the signal amplitude A and phase φ can be summarized by the following equation:

[0083]

[0084] Where ∈ represents the radar cross-section, a commonly used metric for measuring a target's ability to reflect radar signals, P t G t A eff These represent the transmit power, antenna gain, and effective antenna area, respectively, where λ is the wavelength of the radar signal, and d... ra Δd represents the distance information between the target object and the radar antenna. ra It is the change in distance between the target object and the radar caused by the curvature of the object's surface.

[0085] The presence of an object causes the radar's cross-section to change at different observation angles, thus affecting the amplitude. Furthermore, signal scattering caused by minute changes in the curvature of the object's surface can also increase or decrease the phase. Figure 4 This diagram illustrates the variation in the distribution of the intermediate frequency signal reflected from the object in the IQ domain as the incident angle of the radar signal changes. Figure 4 As shown, the clusters of different colors represent the distribution of the intermediate frequency (IF) signal reflected by the object in the IQ domain when the radar signal is incident at different angles. At different incident angles, the surface curvature of the object will cause changes in the amplitude and phase of the signal, resulting in amplitude change ΔA and phase change Δφ. Consequently, the distribution of the IF signal in the IQ domain will exhibit different ring-shaped distributions.

[0086] Figure 5 The experimental results once again proved the above principle. Figure 5 Experimental results are shown for the IQ domain distribution of intermediate frequency signals reflected from objects with different surface curvatures as a function of the incident angle of radar waves. (a) shows the experimental results for concave objects, (b) shows the experimental results for planar objects, (c) shows the experimental results for convex objects, and (d) shows the experimental results for displays.

[0087] exist Figure 5 In this application, four objects with different surface curvatures were selected, and radar was used to detect the objects from multiple incident angles (0°-180°). Each time a signal was received, the object and the radar maintained the same distance. For concave surfaces, the target object tends to concentrate the electromagnetic waves impacting the surface, refocusing the parallel-incident electromagnetic waves at the focal point. Therefore, compared to flat and convex surfaces, concave surfaces have a significant convergence effect on signal amplitude, such as... Figure 5 As shown in (a), for a plane, the distribution of the intermediate frequency (IF) signal in the IQ domain strongly depends on the angle of incidence. Due to specular reflection, a plane has almost no radar cross-section unless directly aligned with the radar (i.e., an angle of incidence of 0° or 180°). Therefore, the distribution of the IF signal of a planar object in the IQ domain will exhibit an extreme phenomenon of dispersion at both ends and concentration at all other angles, such as... Figure 5 As shown in (b). Conversely, for a plane... Figure 5In (c), for convex surfaces, the distribution of the intermediate frequency (IF) signal in the IQ domain is less affected by the incident angle because the radar cross-section of a convex surface rarely varies angularly, especially for spheres or cylinders. Therefore, their IF signals exhibit an almost uniform distribution throughout the IQ domain. Similarly, for complex objects with multiple curvatures, such as displays, the distribution of their IF signals in the IQ domain will show unique differences, such as... Figure 5 As shown in (d).

[0088] To utilize the characteristic that the distribution of the intermediate frequency (IF) signal reflected from an object in the IQ domain changes periodically with the change of the radar signal's incident angle to identify the curvature of the object's surface, this embodiment uses point density as a parameter to characterize the distribution of the IF signal reflected from the object in the IQ domain at a certain angle. In this embodiment, in step S3, the point density characteristics of the IF signal at multiple incident angles are calculated based on the filtered IF signals at multiple incident angles, using the following formula:

[0089]

[0090] Where F is the point density feature of the intermediate frequency signal at each incident angle, N is the number of sampling points of the filtered intermediate frequency signal, and S is the area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain.

[0091] Specifically, the area occupied by the signal clusters of the filtered intermediate frequency signal in the IQ domain can be calculated as follows:

[0092] Figure 6 The diagram illustrates the calculation of the area occupied by a signal cluster in the IQ domain. To facilitate the quantification of the area occupied by signal samples in the IQ domain, a unit area parameter α = 20 is defined to represent the minimum area in the IQ domain. Based on parameter α, each sampling point of the intermediate frequency signal is determined to belong to a specific unit region in the IQ domain. When a sampling point is found to belong to an uncounted unit region, the area is incremented by 1; otherwise, the area remains unchanged.

[0093] Figure 7 The experimental results are shown, illustrating the variation of the point density characteristics of intermediate frequency (IF) signals reflected from objects with different surface curvatures with the radar incident angle. In step S4, the point density characteristics of IF signals at multiple incident angles are sorted from smallest to largest according to the incident angle, thus obtaining the sequence of changes in the point density characteristics of the target object with the incident angle of the radar transmitted signal. Figure 7 As can be seen, the point density features of the embodiments of this application can be used to characterize the distribution of intermediate frequency signals in the IQ domain.

[0094] In one embodiment, a multi-angle LSTM neural network model includes:

[0095] The system consists of a first LSTM layer, a BatchNormalization layer, a second LSTM layer, and a fully connected layer.

[0096] Here, the multi-angle LSTM neural network model is the trained model. During model training, only point density feature data from two consecutive angles are needed. The object's radius of curvature is used as the training label, and the sequence of changes in the object's point density features with the incident angle of the radar signal is used as the training data. The output is passed through a fully connected layer, with class cross-entropy as the loss function of the neural network, and then the RMSprop optimization algorithm is used to obtain the trained multi-angle LSTM neural network model. The specific structure and parameter configuration of the multi-angle LSTM neural network are shown in Table 1.

[0097] Table 1. Specific structure and parameter configuration of the multi-angle LSTM neural network

[0098]

[0099] Employing the same inventive concept as the curvature recognition method based on multi-angle LSTM neural networks, this embodiment also provides a corresponding curvature recognition device based on multi-angle LSTM neural networks. Figure 8 A structural block diagram of a curvature recognition device based on a multi-angle LSTM neural network according to an embodiment of this application is shown. The device includes:

[0100] The intermediate frequency signal acquisition module 81 is used for the radar to transmit signals at multiple incident angles to the target object, the radar to receive the signals reflected by the target object, and to acquire intermediate frequency signals at multiple incident angles based on the signals reflected by the target object.

[0101] The filtering module 82 is used to filter the intermediate frequency signals under multiple incident angles to obtain the filtered intermediate frequency signals under multiple incident angles.

[0102] The point density feature calculation module 83 is used to calculate the point density features of the intermediate frequency signal under multiple incident angles based on the filtered intermediate frequency signal under multiple incident angles.

[0103] The point density feature change sequence acquisition module 84 is used to sort the point density features of intermediate frequency signals under multiple incident angles according to the incident angle from small to large, and obtain the change sequence of the point density features of the target object with the incident angle of the radar transmitted signal.

[0104] The radius of curvature acquisition module 85 is used to input the changing sequence into the multi-angle LSTM neural network model and output the radius of curvature of the target object.

[0105] In one embodiment, the filtering module 82 is further configured to:

[0106] Determine the frequency value of the frequency component corresponding to the target object in the spectrum of the multipath environment;

[0107] The intermediate frequency (IF) signal is subjected to SVMD filtering, which decomposes the IF signal into multiple IMF signals;

[0108] The energy levels of each frequency component contained in each IMF signal are sorted to determine the frequency range in which the main energy of the IMF signal is located.

[0109] Select IMF signals that include the frequency value corresponding to the target object within the frequency range of the main energy source as the filtered IMF signals;

[0110] The filtered intermediate frequency signal is obtained by reconstructing the selected IMF signal.

[0111] In one embodiment, the filtering module 82 is further configured to:

[0112] The distance-energy spectrum is obtained by applying the FFT algorithm to the intermediate frequency signal;

[0113] The CFAR algorithm is used to extract the distance information d between the target object and the radar antenna from the range-energy spectrum. ra ;

[0114] The frequency value is calculated using the following formula: Where K is the slope of the frequency-modulated continuous wave and c is the speed of light.

[0115] In one embodiment, the point density feature calculation module 83 is further configured to calculate the point density feature of the intermediate frequency signal at each incident angle using the following formula:

[0116]

[0117] Where F is the point density feature of the intermediate frequency signal at each incident angle, N is the number of sampling points of the filtered intermediate frequency signal, and S is the area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain.

[0118] The following experiments further verify the effectiveness of the curvature recognition method and device of this application:

[0119] Experimental subjects

[0120] There are 24 kinds of everyday objects, including 12 different curvatures, with the smallest difference in the radius of curvature between the objects being 5mm.

[0121] Experimental environment

[0122] Choose three real-world scenarios: a spacious hall with a few stationary metallic objects; a seminar room with many stationary metallic objects; and an office with a large number of stationary metallic objects.

[0123] Experimental distance

[0124] Choose from 5 distances: 1 meter, 2 meters, 3 meters, 4 meters, and 5 meters from the radar.

[0125] Experimental perspective

[0126] Choose from 5 combinations of incident angles: 0° and 20°, 40° and 60°, 90° and 120°, 140° and 160°, and 160° and 180°.

[0127] Radar movement method in the experiment

[0128] Choose from three movement methods for transmitting and receiving radar signals: stationary, handheld, and pushed by a cart.

[0129] Experimental equipment

[0130] The experiment was conducted on a commercial millimeter-wave radar device, specifically a Texas Instruments IWR1843Boost and DCA1000EVM, capable of transmitting chirp signals from 77 GHz to 81 GHz. It features three TX transmit antennas and four RX receive antennas. To ensure minimum range resolution (3.75 cm), one TX antenna transmits an FMCW signal starting at 77 GHz with a bandwidth of 3.99 GHz. All RX antennas receive reflected signals. Each FMCW signal is modulated with the following parameters: chirp slope = 20 MHz / μs; baseband sampling rate = 3 Msps; ADC sample count = 512; maximum antenna gain = 10 dBi.

[0131] Experiment 1

[0132] The objective of Experiment 1 was to verify the impact of different environments on the curvature recognition rate of this application. The distance between the object and the radar was maintained at 1 meter, the radar incident angle combination was 0 degrees and 20 degrees, and the object remained stationary while transmitting and receiving signals from the radar.

[0133] Test results of Experiment 1:

[0134] Figure 9 The following figures illustrate the experimental results of curvature recognition using the method described in this application under different environments: Figure 9 As shown, the curvature recognition accuracy of this application gradually decreases as the environment becomes more complex, but it always maintains a recognition rate of over 98%, which proves that the curvature recognition method of this application can maintain a high curvature recognition rate in both simple and complex environments.

[0135] Experiment 2

[0136] The purpose of Experiment 2 was to verify the effect of different distances on the curvature recognition rate of this application. The experimental site was fixed in the office, with radar incident angle combinations of 0 degrees and 20 degrees, and the radar remained stationary during signal transmission and reception.

[0137] Test results of Experiment 2:

[0138] Figure 10 The following figures illustrate the experimental results of curvature recognition using the method described in this application at different distances: Figure 10 As shown, the curvature recognition accuracy of this application fluctuates as the distance between the object and the radar increases, but it always maintains a recognition rate of over 94%, which proves that the curvature recognition method of this application can maintain a high curvature recognition rate in both close-range and long-range indoor environments.

[0139] Experiment 3

[0140] The purpose of Experiment 3 was to verify the effect of different incident angles of radar signals on the curvature recognition rate of this application. The experimental site was fixed in the office, the experimental distance was fixed at 1 meter, and the radar remained stationary during signal transmission and reception.

[0141] Test results of Experiment 3:

[0142] Figure 11 The following diagram shows the experimental results of curvature recognition using the method of this application at different angles: Figure 11 As shown, the curvature recognition accuracy of this application fluctuates with the change of the radar signal incident angle combination, but it always maintains a recognition rate of over 97%, which proves that the curvature recognition method of this application can maintain a high curvature recognition rate under multiple angle combinations in complex indoor close-range environments.

[0143] Experiment 4

[0144] The purpose of Experiment 4 was to verify the effect of different motion states during radar signal transmission and reception on the curvature recognition rate of this application. The experimental site was fixed in the office, the experimental distance was fixed at 1 meter, and the radar incident angle combination was 0 degrees and 20 degrees.

[0145] Test results of Experiment 4:

[0146] Figure 12 The following diagram shows the experimental results of curvature recognition using the method described in this application under different radar motion states: Figure 12 As shown, the curvature recognition accuracy of this application fluctuates with the change of the motion state when the radar transmits and receives signals, but it always maintains a recognition rate of over 94%. This proves that the curvature recognition method of this application can maintain a high curvature recognition rate in various radar motion states in complex indoor close-range environments.

[0147] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A curvature recognition method based on a multi-angle LSTM neural network, characterized in that, The method comprises the following steps: The radar transmits signals of multiple incident angles to a target object, the radar receives signals reflected by the target object, and intermediate frequency signals under multiple incident angles are obtained according to the signals reflected by the target object; The intermediate frequency signals under multiple incident angles are filtered to obtain filtered intermediate frequency signals under multiple incident angles; Point density characteristics of the intermediate frequency signals under multiple incident angles are calculated according to the filtered intermediate frequency signals under multiple incident angles; The point density characteristics of the intermediate frequency signals under multiple incident angles are sorted in ascending order of incident angles to obtain a sequence of changes of the point density characteristics of the target object with the incident angles of radar transmitted signals; The sequence is input into a multi-angle LSTM neural network model to output a curvature radius of the target object; The filtering of the intermediate frequency signals under multiple incident angles to obtain the filtered intermediate frequency signals under multiple incident angles comprises the following steps: The frequency value of the frequency component corresponding to the target object in the frequency spectrum of the multipath environment is determined; The intermediate frequency signals are filtered by SVMD to decompose the intermediate frequency signals into multiple IMF signals; The energy size of each frequency component contained in each IMF signal is sorted to determine the frequency range in which the main energy of the IMF signal is located; An IMF signal in which the frequency value corresponding to the target object is included in the frequency range in which the main energy is located is selected as a screened IMF signal; The filtered intermediate frequency signals are reconstructed according to the screened IMF signal; The multi-angle LSTM neural network model comprises: A first LSTM layer, a BatchNormalization layer, a second LSTM layer and a fully connected layer.

2. The method of claim 1, wherein, The determination of the frequency value of the frequency component corresponding to the target object in the frequency spectrum of the multipath environment comprises the following steps: The distance-energy spectrum is obtained by using the FFT algorithm on the intermediate frequency signals; The point density characteristics of the intermediate frequency signals under multiple incident angles are calculated according to the filtered intermediate frequency signals under multiple incident angles by using the following formula: A CFAR algorithm is applied to the distance-energy spectrum to extract distance information between the target object and the radar antenna ; The frequency value is calculated using the following equation: wherein, is the slope of the frequency modulated continuous wave, is the speed of light.

3. The method of claim 1, wherein, The method comprises the following steps: = wherein, is a point density feature of the intermediate frequency signal at each incident angle, is a number of sampling points of the filtered intermediate frequency signal, is an area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain. 4.A curvature recognition device based on a multi-angle LSTM neural network, characterized in that, An intermediate frequency signal acquisition module is configured to transmit signals of multiple incident angles to a target object by a radar, receive signals reflected by the target object by the radar, and obtain intermediate frequency signals under multiple incident angles according to the signals reflected by the target object; A filtering module is configured to filter the intermediate frequency signals under multiple incident angles to obtain filtered intermediate frequency signals under multiple incident angles; A point density characteristic calculation module is configured to calculate point density characteristics of the intermediate frequency signals under multiple incident angles according to the filtered intermediate frequency signals under multiple incident angles; A point density characteristic change sequence acquisition module is configured to sort the point density characteristics of the intermediate frequency signals under multiple incident angles in ascending order of incident angles to obtain a sequence of changes of the point density characteristics of the target object with the incident angles of radar transmitted signals; A curvature radius acquisition module is configured to input the sequence into a multi-angle LSTM neural network model to output a curvature radius of the target object; The filtering module is further configured to: ​ Determine a frequency value of a frequency component corresponding to the target object in a spectrum of a multipath environment; Perform SVMD filtering processing on the intermediate frequency signal to decompose the intermediate frequency signal into a plurality of IMF signals; Sort energy sizes of frequency components contained in each IMF signal to determine a frequency range in which main energy of the IMF signal is located; Select an IMF signal in which the frequency value corresponding to the target object is included in the frequency range in which the main energy is located as a screened IMF signal; Reconstruct a filtered intermediate frequency signal according to the screened IMF signal; The multi-angle LSTM neural network model comprises: a first LSTM layer, a BatchNormalization layer, a second LSTM layer and a full connection layer.

5. The apparatus of claim 4, wherein, The filtering module is further configured to: obtain a distance-energy spectrum by using an FFT algorithm on the intermediate frequency signal; A CFAR algorithm is applied to the distance-energy spectrum to extract distance information between the target object and the radar antenna ; The frequency value is calculated using the following equation: wherein, is the slope of the frequency modulated continuous wave, is the speed of light.

6. The apparatus of claim 4, wherein, The point density feature calculation module is further configured to calculate a point density feature of the intermediate frequency signal under each incident angle by using the following formula: = wherein, is a point density feature of the intermediate frequency signal at each incident angle, is a number of sampling points of the filtered intermediate frequency signal, is an area occupied by the signal cluster of the filtered intermediate frequency signal in the IQ domain.

Images