Object recognition method and object recognition device

Abstract

An object recognition method includes generating Doppler spectrum data based on an echo signal, wherein the echo signal is related to the object; converting N sets of time domain data corresponding to N velocities in the Doppler spectrum data into N sets of spectrum data; combining the N sets of spectrum data to obtain gait spectrum data; and obtaining gait features from the gait spectrum data to identify the object.

Description

Technical Field

[0001] The present invention relates to an object recognition method, and more particularly to an object recognition method and an object recognition device using radar. Background Art

[0002] Target recognition is a crucial technology for classifying people, animals, vehicles, and other moving objects. It is widely used in many fields, such as surveillance and intrusion detection. Among various object recognition technologies, radar has gained attention due to its robustness to nighttime, harsh environments, and poor lighting conditions. Radar can detect the size of targets within its range and its speed based on the frequency shift between the transmitted and reflected waves using the Doppler effect. However, distinguishing targets solely based on size and speed has many limitations, hindering the accuracy and speed of target recognition. Summary of the Invention

[0003] An embodiment provides an object recognition method, comprising generating Doppler spectrum data based on an echo signal, wherein the echo signal is related to the object; converting N sets of time domain data corresponding to N speeds on the Doppler spectrum data into N sets of spectrum data respectively; combining the N sets of spectrum data to obtain gait spectrum data; and obtaining gait features from the gait spectrum data to identify the object.

[0004] An embodiment provides another object recognition device comprising a radar and a processor. The radar is configured to receive an echo signal associated with an object. The processor is coupled to the radar and configured to generate Doppler spectrum data based on the echo signal, convert N sets of time-domain data corresponding to N velocities in the Doppler spectrum data into N sets of spectrum data, combine the N sets of spectrum data to obtain gait spectrum data, and extract gait features from the gait spectrum data to identify the object. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1 2 is a block diagram of an object recognition device in an embodiment of the present invention.

[0006] Figure 2 for Figure 1 Flowchart of an object recognition method used by an object recognition device.

[0007] Figure 3 Schematic diagram showing preprocessed data.

[0008] Figure 4 A Doppler spectrum diagram is shown in an embodiment of the present invention.

[0009] Figure 5 A Fourier transform spectrum diagram is shown in an embodiment of the present invention.

[0010] Figure 6Shows a pace spectrum diagram according to an embodiment of the present invention.

[0011] Figure 7 A one-dimensional pace spectrum diagram containing positive speed pace spectrum data according to an embodiment of the present invention is shown.

[0012] Figure 8 A one-dimensional pace spectrum diagram including positive speed pace spectrum data and negative speed pace spectrum data is shown in an embodiment of the present invention.

[0013] Figure 9 The combined Doppler spectrum diagram in an embodiment of the present invention is shown.

[0014] Figure 10 Displays the autocorrelation values generated from the combined paced spectrum data.

[0015] Figure 11 Shows the autocorrelation of the combined pace spectrum data according to an embodiment of the present invention.

[0016] Figure 12 Shown is a velocity-normalized Doppler spectrum diagram according to an embodiment of the present invention.

[0017] Explanation of symbols

[0018] 1: Object recognition device

[0019] 10: Radar

[0020] 12: Processor

[0021] 120: Classifier

[0022] 200: Object Recognition Method

[0023] S202 to S214: Steps

[0024] AC(r),AC(r+1):autocorrelation function

[0025] Ds: sampling data

[0026] Ea: Normalized total energy

[0027] Ec: Combined Energy

[0028] fc: pacing frequency

[0029] h1 to h8: speed index value

[0030] L: Window length

[0031] L1 to L8: Sampling lines

[0032] M: total number of time windows

[0033] O: Object

[0034] p: time index value

[0035] s[0,h1] to s[39,h8]: time domain data

[0036] S[0,h1] to S[99,h1]: spectrum data

[0037] Sd: Doppler signal

[0038] Se: echo signal

[0039] St: Transmit signal

[0040] t: time

[0041] T(0) to T(M-1): time interval

[0042] v, v1, v2: speed

[0043] vn: normalized speed

[0044] f0, fm, f2, f3: peak frequencies

[0045] x[0] to x[ML-1]: pre-processed data

[0046] r: displacement DETAILED DESCRIPTION

[0047] Figure 1 The block diagram of the object recognition device 1 according to an embodiment of the present invention is shown. The object recognition device 1 can recognize an object O. The recognized object O can be a person, cat, dog, animal, or other moving object. Because the walking posture of different people or animals, or the way objects move, has its own specific pattern, also known as gait characteristics, the object recognition device 1 can recognize the object O based on the gait characteristics of the object O. Gait characteristics can include the speed and direction of the torso, the speed and direction of the swinging of the limbs, the speed and direction of the swinging of other attached objects, the flapping of bird wings, the rotation of a fan, the vibration of an engine, etc. The object recognition device 1 can include a radar 10, an analog-to-digital converter (ADC) 11, and a processor 12. The radar 10 is coupled to the processor 12 via the ADC 11. After sending a transmission signal St, the radar 10 can detect an echo signal Se related to the movement of the object O. The processor 12 may process the echo signal Se to obtain the gait characteristics of the object O, and identify the object O in a classification manner according to the gait characteristics.

[0048] Radar 10 can be a continuous wave radar, a frequency modulated continuous wave radar, or another type of radar. Radar 10 can emit a transmission signal St and receive an echo signal Se when it detects an object O. Transmission signal St has a predetermined frequency. When object O is in motion, the relative radial motion between object O and radar 10 causes a change in the frequency of echo signal Se, known as Doppler shift. Because radars can be fixed in position and the Doppler shift is related to the speed of relative motion, the Doppler shift can be used to estimate the velocity of object O when it is in motion. Equation 1 is as follows:

[0049]

[0050] Where v is the velocity of object O;

[0051] c is the speed of light;

[0052] ft is the frequency of the transmitted signal St; and

[0053] fd is the frequency of the echo signal Se minus ft, which is called the Doppler frequency.

[0054] When object O moves directly toward radar 10, its velocity v is proportional to the Doppler frequency fd. As object O approaches radar 10, its velocity v is positive and its Doppler frequency fd is positive. As object O moves away from radar 10, its velocity v is negative, and both its Doppler frequency fd and velocity v are negative. The echo signal Se may have at least one frequency. In some embodiments, radar 10 may mix the echo signal Se with the transmitted signal St to generate a Doppler signal Sd having at least one Doppler frequency fd. ADC 11 may sample Doppler signal Sd using a specific sampling frequency, such as 44100 Hz, to generate a plurality of sampled data Ds. The processor 12 may pre-process the plurality of sampled data Ds to generate a plurality of pre-processed data. The pre-processed data are then divided into M segments for short-time-interval frequency domain transformation, sequentially generating M sets of frequency energy distributions. Each set of frequency energy distributions corresponds to the frequency domain range of a Doppler frequency fd and forms an energy spectrum, where M is a positive integer. In this embodiment, the M segments of pre-processed data may partially overlap. Because at least one Doppler frequency fd in the Doppler signal Sd can be converted into an energy distribution of M sets of frequencies, the processor 12 may use Equation 1 to convert the sampling results of the M sets of frequencies in the Doppler signal Sd into M sets of velocities. After the conversion according to Equation 1, the M sets of velocity data can be considered as M sets of Doppler spectrum data for the Doppler signal Sd at M time intervals. The processor 12 may further perform a discrete frequency transformation on the M sets of Doppler spectrum data to generate cadence spectrum data, and extract gait characteristics of the object O from the cadence spectrum data. Discrete frequency transformation can be implemented using fast Fourier transform. The processor 12 may include a classifier 120 to identify the object O based on gait characteristics. The classifier 120 can be implemented using a support vector machine (SVM) algorithm, a K-nearest neighbors (KNN) algorithm, a linear discriminant analysis algorithm, other classification algorithms, or a combination thereof. Short-time frequency domain transformation can be implemented by a short-time Fourier transform, a wavelet transform, a Hilbert-Huang transform, or a combination thereof. Discrete frequency transformation can be a discrete Fourier transform or a fast Fourier transform. In some embodiments, an independent time-frequency transformation circuit can be used to perform short-time frequency transformation on M segments of pre-processed data, and / or an independent discrete frequency transformation circuit can be used to convert M groups of Doppler spectrum data into cadence spectrum data.

[0055] Figure 2 The flowchart of the object recognition method 200 used by the object recognition device 1 includes steps S202 to S214, wherein step S202 is used to detect the object O, and steps S204 to S214 are used to identify the object O. Any reasonable technical changes or step adjustments fall within the scope of the present invention. The following describes steps 202 to S214 in detail using the object recognition device 1:

[0056] Step S202: When the object O is detected, the radar 10 receives the echo signal Se;

[0057] Step S204: the processor 12 pre-processes the echo signal Se to obtain pre-processed data;

[0058] Step S206: The processor 12 converts the pre-processed data into M sets of Doppler spectrum data;

[0059] Step S208: The processor 12 converts the N groups of time domain data corresponding to the N speeds in the M groups of Doppler spectrum data into N groups of spectrum data respectively;

[0060] Step S210: The processor 12 combines N sets of spectrum data to obtain pace spectrum data;

[0061] Step S212: the processor 12 obtains gait features from the gait spectrum data;

[0062] Step S214: The processor 12 identifies the object O according to the gait characteristics.

[0063] In step S202, the radar 10 continuously transmits a transmit signal St. When an object O is detected, the radar 10 receives an echo signal Se and generates a Doppler signal Sd based on the echo signal Se. In step S204, the ADC 11 samples the Doppler signal Sd to generate a plurality of sampled data Ds. The processor 12 then pre-processes the plurality of sampled data Ds to obtain a plurality of pre-processed data. Pre-processing may include reducing the sampling frequency, removing signal interference, and increasing resolution, and can be implemented using software, hardware, or a combination thereof. The processor 12 may reduce the number of sampled data Ds, for example, by using a resampling function to downsample 44,100 sampled data Ds per second by a factor of 80 to generate approximately 550 resampled data per second. Resampling data reduces the computational complexity of subsequent signal processing and prevents filtering of the sampled data Ds from being too large to be processed by the filter, which could cause signal distortion and affect the extraction of gait features. Removing signal interference may include removing mains signal interference and / or removing short-time window signal processing interference. Next, a notch filter reduces or removes AC interference from the resampled data to generate reduced-interference data. For example, since the frequency of a typical mains AC signal is 60 Hz, the notch filter can reduce or remove harmonic signal interference in frequency bands of 60 Hz and its multiples, such as 120 Hz, 180 Hz, and 240 Hz, from the resampled data. It should be understood that to avoid energy leakage, the spectrum of the sampled data Ds of the Doppler signal Sd can be generated using M short-time frequency domain transforms. These M short-time frequency domain transforms use a Hamming window to perform window-type masking on the signal to reduce data interference, suppress the energy of the spectrum on both sides, and maximize the energy of the spectrum main lobe. The window length of the Hamming window can be the same as the window length of the window function used in the short-time frequency transform, for example, the same as the window length of the short-time Fourier transform, wavelet transform, or Hilbert-Huang transform. In some embodiments, the order of removing mains signal interference and removing short-time window signal processing interference can be interchanged. Finally, processor 12 increases the number of filtered data sampling points, for example, by four times the number of filtered data sampling points, from 550 sampling points per second to 2200 sampling points per second. The increase in the number of pre-processed data sampling points also increases the resolution required for short-time frequency conversion.

[0064] In step S206 , the processor 12 uses short-time Fourier transform to generate M sets of Doppler spectrum data according to the plurality of pre-processed data. Figure 3Schematic diagram showing ML pre - processed data x[0] to x[ML - 1], where the horizontal axis represents the time index value p, and the vertical axis represents the energy intensity of the pre - processed data x[p], for example, voltage. M and L are positive integers, p is an integer and 0 ≤ p < ML. When performing the short - time Fourier transform, the processor 12 can first divide the ML pre - processed data into processed data for M time intervals T(0) to T(M - 1), and the processed data for each time interval corresponds to L sampling points. The processor 12 can then use a window function w(·) with a fixed length L to process the sampling points for each time interval. The first time interval T(0) corresponds to the pre - processed data x[0] to x[L - 1]. The second time interval T(1) corresponds to the pre - processed data x[L] to x[2L - 1], and so on. The m - th time interval T(M - 1) corresponds to the pre - processed data x[(m - 1)L] to x[mL - 1]. m is a non - negative integer and m ≤ M. The processor 12 can perform the Fourier transform on the L pre - processed data for each time interval to generate the energy distribution of M groups of frequencies. In some embodiments, two adjacent time intervals can partially overlap. For example, the pre - processed data within 4 seconds is divided into time intervals with a time length of 0.1 seconds to generate 40 time intervals. Each time interval corresponds to 220 sampling points of the pre - processed data. The processor 12 can perform the Fourier transform on the 220 sampling points of the pre - processed data corresponding to each time interval to generate the energy distribution of the frequencies for 40 time intervals. In some embodiments, for the convenience of hardware operation, the processor 12 can supplement some data points with a value of 0 in the data points within each time interval to obtain 2 to the power of N frequency energies within each time interval, for example, 256 frequency energies, where 256 frequency energies include 128 positive frequency energies and 128 negative frequency energies. The short - time Fourier transform can be a discrete short - time Fourier transform, which is represented by Equation 2:

[0065]

[0066] where L is the window length;

[0067] w[q] is the window function data, q is the window function data index value, q is an integer and 0 ≤ q < L;

[0068] x[(m - 1)L+q] is the pre - processed data for the m - th time interval T(m - 1), m is the time interval index value; and

[0069] X[k,m,h] is the energy of the m - th group of frequencies, k is the time index value, k is an integer and 0 ≤ k < L, h is the frequency index value, h is an integer and 0 ≤ h < L.

[0070] Each frequency index value h corresponds to a frequency fd, and the processor 12 can use Formula 1 to convert the frequency fd into a corresponding velocity v. In some embodiments, when 0 ≤ h ≤ (L / 2-1), the frequency fd can correspond to a positive velocity v; when L / 2 ≤ h ≤ (L-1), the frequency fd can correspond to a negative velocity v. The processor 12 can then generate the energy s[m,h] corresponding to the velocity v based on the amplitude of the energy X[k,m,h] of the mth frequency group, as expressed by Formula 3:

[0071] s[m,h]=|X[k,m,h]| 2 Formula 3

[0072] The processor 12 can generate M sets of Doppler spectrum data based on the energy s[m,h] of velocity v at all times. The mth set of Doppler spectrum data includes the energy s[m,0] to s[m,L-1] of L velocities v. The M sets of Doppler spectrum data can form a Doppler spectrum matrix Dm, which is represented by Equation 4:

[0073]

[0074] When m is 0 to M-1 and h is L / 2-1 to 0, the matrix element s[m,h] of the Doppler spectrum matrix Dm is the energy corresponding to the positive velocity v. When m is 0 to M-1 and h is L / 2 to L-1, the matrix element s[m,h] of the Doppler spectrum matrix Dm is the energy corresponding to the negative velocity v. The processor 12 can generate data corresponding to the Doppler spectrogram based on the Doppler spectrum matrix Dm, such as Figure 4 As shown in the figure, the horizontal axis represents time t and the vertical axis represents velocity v. The grayscale in the figure represents velocity energy, with higher grayscale areas indicating higher energy s[m,h]. The vertical axis at a given time t represents the energy distribution of that velocity. For example, when time t is 0, the energy distribution corresponds to the first group of velocities ranging from 4m / s to -2m / s. The velocity set for each time t can contain 128 positive and 128 negative velocities, corresponding to the back-and-forth swinging of the torso and limbs of a person or animal when moving.

[0075] In step S208, the processor 12 converts the N sets of time domain data of the N speeds v among the L speeds on the M sets of Doppler spectrum data into N sets of spectrum data. Figure 4Four sampling lines L1, L2, L3, and L4 (N = 4) are drawn on the Doppler spectrogram in , corresponding to four velocities of 1 m / s, 2 m / s, 3 m / s, and 4 m / s respectively. The processor 12 obtains the first set of time-domain data {s[0, h1], s[1, h1], …, s[39, h1]} corresponding to the velocity of 1 m / s from the sampling line L1. The second set of time-domain data {s[0, h2], s[1, h2], …, s[39, h2]} corresponding to the velocity of 2 m / s is obtained from the sampling line L2. The third set of time-domain data {s[0, h3], s[1, h3], …, s[39, h3]} corresponding to the velocity of 3 m / s is obtained from the sampling line L3. The fourth set of time-domain data {s[0, h4], s[1, h4], …, s[39, h4]} corresponding to the velocity of 4 m / s is obtained from the sampling line L4. The time-domain data s[m, hn] represents the energy at the sampling line Ln and the m-th time window, where n is a positive integer and n < N. The processor 12 performs a discrete Fourier transform on the n-th set of time-domain data {s[0, hn], s[1, hn], …, s[39, hn]} using Equation 5 to generate the n-th set of spectral data {S[0, hn], S[1, hn], …, S[M - 1, hn]}.

[0076]

[0077] Where S[f, hn] is the energy of the n-th set of spectral data at the (f + 1)-th frequency, f is a non-negative integer less than M, and M is the number of samples. For example, the processor 12 can perform a discrete Fourier transform on the first set of time-domain data {s[0, h1], s[1, h1], …, s[39, h1]} to obtain the first set of spectral data {S[0, h1], S[1, h1], …, S[39, h1]}, perform a discrete Fourier transform on the second set of time-domain data {s[0, h2], s[1, h2], …, s[39, h2]} to obtain the second set of spectral data {S[0, h2], S[1, h2], …, S[39, h2]}, perform a discrete Fourier transform on the third set of time-domain data {s[0, h3], s[1, h3], …, s[39, h3]} to obtain the third set of spectral data {S[0, h3], S[1, h3], …, S[39, h3]}, and perform a discrete Fourier transform on the fourth set of time-domain data {s[0, h4], s[1, h4], …, s[39, h4]} to obtain the fourth set of spectral data {S[0, h4], S[1, h4], …, S[39, h4]}. The first set of spectral data {S[0, h1], S[1, h1], …, S[39, h1]} can generate a spectrogram corresponding to the velocity of 1 m / s according to the frequency, such as Figure 5As shown, the horizontal axis represents frequency f and the vertical axis represents energy E. The frequency spectrum of a speed of 1m / s represents the energy distribution at different frequencies, which can represent the number of swings per second of a specific part of the object O. Figure 5 The main frequency distribution of the display speed of 1m / s is less than 10Hz. The spectrum data of other speeds of 2m / s, 3m / s, and 4m / s can also be drawn into a spectrum diagram according to the frequency. In some embodiments, taking into account the normal swing frequency caused by people or animals walking, the frequency range considered by the spectrum diagram can be set to 0-20Hz. In some embodiments, the processor 12 can also generate spectrum data of negative velocity based on the time domain data corresponding to the negative velocity on the Doppler spectrum data, for example, based on the four groups of time domain data {s[0,h5],s[1,h5],…,s[39,h5]} to {s[0,h8],s[1,h8],…,s[39,h8]} on the sampling lines L5 to L8 to generate four groups of negative velocity spectrum data {S[0,h5],S[1,h5],…,S[39,h5]} to {S[0,h8],S[1,h8],…,S[39,h8]}.

[0078] In step S210, the processor 12 combines N sets of spectral data to obtain paced spectral data. In some embodiments, the processor 12 may combine the N sets of spectral data along the vertical axis to obtain the paced spectral data. For example, the processor 12 may combine the energies of the corresponding frequency domains in the first set of spectral data {S[0,h1], S[1,h1],…, S[39,h1]} to the eighth set of spectral data {S[0,h8], S[1,h8],…, S[39,h8]} along the vertical axis to generate a matrix {C[0],…, C

[39] }. C[0] to C

[39] are row vectors of the matrix. Each matrix row vector C[0] to C

[39] is a set of cadence spectrum data, C[0] = g(0)×(S[0,h1]; S[0,h2]; ...; S[0,h8]), C

[39] = g(39)×(S[39,h1]; S[39,h2]; ...; S[39,h8]), where g is a normalization coefficient that can be given different weights according to different corresponding speeds. Each set of cadence spectrum data C[0] to C

[39] represents the normalized speed energy distribution at the corresponding frequency. The matrix {C[0], ..., C

[39] } can generate a corresponding cadence spectrogram. Figure 6 The graph is a step spectrum diagram, where the horizontal axis represents the step frequency fc. The vertical axis represents the speed. The grayscale in the graph represents the energy corresponding to the speed. The higher the grayscale, the higher the energy. The step spectrum diagram represents the energy distribution corresponding to the speed under different step frequencies fc. For example, Figure 6In the example, the energy distribution corresponding to the velocity of object O at cadence frequencies fc of 0 Hz, 2 Hz, 4 Hz, and 6 Hz can be observed. 0 Hz may be the swing frequency of the trunk, 2 Hz may be the swing frequency of the limbs, and 4 Hz and 6 Hz may be harmonics of 2 Hz. In some embodiments, processor 12 may remove the DC component of the cadence spectrum data, i.e., the cadence spectrum data C[0] located at approximately 0 Hz.

[0079] In other embodiments, the processor 12 may sum N sets of spectral data along the vertical axis to obtain pace spectrum data. The N sets of spectral data may correspond to positive speed and / or negative speed. For example, when using N sets of spectral data corresponding to positive speed, the processor 12 may sum the first set of spectral data {S[0,h1], S[1,h1],…, S[39,h1]} to the fourth set of spectral data {S[0,h4], S[1,h4],…, S[39,h4]} element-wise along the vertical axis to generate vectors C[0],…, C

[39] . Each vector C[0] to C

[39] corresponds to a set of pace spectrum data. C[0] = g×(S[0,h1] + S[0,h2] +…+S[0,h4]). C

[39] = g×(S[39,h1] + S[39,h2] +…+S[39,h4]). Each vector C[0] to C

[39] represents the normalized velocity sum energy of the corresponding frequency. Vectors C[0] to C

[39] can be summed to generate a one-dimensional cadence spectrum. Figure 7 is a one-dimensional step spectrum, where the horizontal axis represents the step frequency fc and the vertical axis represents the normalized total energy Ea of the vector C. The one-dimensional step spectrum represents the normalized total energy distribution at different step frequencies fc. Figure 7 In Figure 1, the normalized total energy Ea distribution corresponding to the velocity of object O at cadence frequencies fc of 0 Hz, 2 Hz, 4 Hz, and 6 Hz can be observed. The normalized total energy Ea peak at 0 Hz can be the swing frequency of the trunk, 2 Hz can be the swing frequency of the limbs, and 4 Hz and 6 Hz can be harmonics of 2 Hz. In some embodiments, processor 12 can remove the DC component of the cadence spectrum data, that is, the cadence spectrum data C[0] at approximately 0 Hz.

[0080] The processor 12 then obtains gait features from the gait spectrum data (step S212) and identifies the object O based on the gait features (step S214). The gait features may include: 1. the ratio of movement amplitude and secondary component energy to primary component energy in the gait spectrum graph; 2. gait spectrum data from a one-dimensional gait spectrum graph; 3. gait spectrum data and autocorrelation minimum value from a merged gait spectrum graph; or 4. velocity-normalized gait spectrum data.

[0081] The processor 12 may use the stride and the ratio of the secondary component energy to the primary component energy in the stride spectrum as gait features to identify the object O. The object O includes a primary component and a secondary component. For example, the primary component may be the human torso, and the secondary components may be the human limbs. The processor 12 may identify the fundamental frequency and the primary velocity of the primary component from the stride spectrum data, generate the stride based on the fundamental frequency and the primary velocity, identify the secondary component energy corresponding to the secondary component and the primary component energy corresponding to the primary component from the stride spectrum data, generate the ratio of the secondary component energy to the primary component energy, and identify the object O based on the stride and the ratio. The primary component velocity may be the torso velocity of the torso. The stride may be the stride length S. The primary component energy may be the portion of the echo signal Se corresponding to the torso energy. The secondary component energy may be the portion of the echo signal Se corresponding to the limb energy. The ratio of the secondary component energy to the primary component energy may be the ratio R of the limb energy to the torso energy. In some embodiments, the processor 12 may use step length S and the ratio of limb energy to trunk energy R as gait features to identify the object O. The processor 12 may use Equation 6 to calculate the step length S based on the trunk velocity and fundamental frequency, and use Equation 7 to calculate the ratio of limb energy to trunk energy R to generate the gait features.

[0082]

[0083] Where S is the stride length;

[0084] Vt is the trunk velocity; and

[0085] fm is the fundamental frequency.

[0086]

[0087] Where R is the limb-to-trunk ratio;

[0088] n is the index value of the secondary part energy;

[0089] D is the maximum index value of the secondary part energy;

[0090] S0 is the main part of energy; and

[0091] Sn is the minor part of energy.

[0092] refer to Figure 6, the corresponding maximum longitudinal velocity v1 or v2 across the stride frequency of 0 Hz to 20 Hz can be the trunk velocity Vt. The stride frequency fm corresponding to the longitudinal velocity of 7 m / s to -5 m / s can be the fundamental frequency of the limb swing. For example, the trunk velocity Vt is approximately 2 m / s, the fundamental frequency fm is approximately 2 Hz, and the stride length S is approximately 1 meter (Vt / fm=2 / 2=1). In some embodiments, the processor 12 can add the energy of the same velocity v along the horizontal axis to generate a total energy, and determine the corresponding speed of the maximum total energy among all the total energies as the trunk velocity Vt, for example, 2 m / s. The processor 12 can add the energy of the same frequency f along the vertical axis to generate a total energy, and determine the peak frequency based on all the total energies, and then determine the maximum peak frequency of the non-DC frequency as the fundamental frequency fm, for example, 2 Hz.

[0093] Furthermore, the energy distribution at peak frequencies f0, fm, f2, and f3 on the horizontal axis may correspond to energy distributions spanning speeds of 7 m / s to -7 m / s, 7 m / s to -5 m / s, 7 m / s to -5 m / s, and 6 m / s to -4 m / s. For example, the primary energy component S0 may be the summed energy corresponding to speeds of 7 m / s to -7 m / s at the peak frequency f0, and the maximum index value D may be 3. The secondary energy component may include first to third secondary energy components S1 to S3, where the first secondary energy component S1 may be the summed energy corresponding to speeds of 7 m / s to -5 m / s at the peak frequency fm, the second secondary energy component S2 may be the summed energy corresponding to speeds of 7 m / s to -5 m / s at the peak frequency f2, and the third secondary energy component S3 may be the summed energy corresponding to speeds of 6 m / s to -4 m / s at the peak frequency f3.

[0094] The ratio R of the limbs to the trunk may be (S1+S2+S3) / S0. In some embodiments, the processor 12 may add the energy at each frequency f along the longitudinal axis to generate the total energy at that frequency. After the processor 12 obtains the total energy at all frequencies, it may determine the peak frequency. The processor 12 may define the peak frequency approximately equal to 0 Hz as the peak frequency f0. The processor 12 may define the peaks after 0 Hz as fm, f2, and f3 in sequence. The processor 12 may further define the total energy corresponding to the frequencies f0, fm, f2, and f3 as the main energy component S0 and the secondary energy components S1, S2, and S3 in sequence.

[0095] Processor 12 may input stride length S and limb-to-torso ratio R into classifier 120 to classify object O based on stride length S and limb-to-torso ratio R. Different objects O may have different stride lengths S. For example, a walking person's stride length S may be between 30 and 40 centimeters, while a walking small dog's stride length S may be less than 10 centimeters. Different objects O may also have different limb-to-torso ratios R. For example, a human's limb-to-torso ratio R may be approximately 0.6, while a dog's limb-to-torso ratio R may be greater than 0.7.

[0096] The processor 12 may use the one-dimensional gait spectrum data as a gait feature to identify the object O. Figure 7 , the processor 12 may input the pace spectrum data C[0] to C

[39] to the classifier 120 to classify the object O according to the pace spectrum data C[0] to C

[39] . In some embodiments, the processor 12 may convert the energy distribution of the positive velocity and the energy distribution of the negative velocity of the Doppler spectrum into positive velocity pace spectrum data and negative velocity pace spectrum data, respectively, for example Figure 5 The four sets of time domain data on the sampling lines L1 to L4 are converted into positive speed step spectrum data C[0] to C

[39] , and the four sets of time domain data on the sampling lines L5 to L8 are converted into negative speed step spectrum data C

[40] to C

[79] . The positive speed step spectrum data C[0] to C

[39] and the negative speed step spectrum data C

[40] to C

[79] can be integrated into a one-dimensional step spectrum, such as Figure 8 As shown, the horizontal axis represents the pace frequency fc, and the vertical axis represents the normalized total energy Ea of the pace spectrum data C. The processor 12 can input the pace spectrum data C[0] to C

[79] into the classifier 120 to classify the object O according to the pace spectrum data C[0] to C

[79] . Different objects O may correspond to different peak pace frequencies. For example, the peak pace frequency of a person may be at a position of about 2 Hz, and its higher harmonics may be at a position of about 4 Hz and 6 Hz, and the peak pace frequency of a dog may be unfixed. In some embodiments, the processor 12 may put the pace spectrum data of the larger one of the total energy of the positive speed pace spectrum data C[0] to C

[39] and the total energy of the negative speed pace spectrum data C

[40] to C

[79] before the pace spectrum data of the other as gait features and input them into the classifier 120. The larger one of the total energy in the positive speed pace spectrum data and the negative speed pace spectrum data can be expressed as:

[0097] max{Ea(C[0] to C

[39] ),Ea(C

[40] to C

[79] )}

[0098] For example, when the total energy of the negative speed step spectrum data C

[40] to C

[79] is greater than the total energy of the positive speed step spectrum data C[0] to C

[39] , the gait feature is:

[0099] {C

[40] ,…,C

[79] ,C[0],…,C

[39] }

[0100] When the total energy of the positive speed step spectrum data C[0] to C

[39] is greater than the total energy of the negative speed step spectrum data C

[40] to C

[79] , the gait feature is:

[0101] {C[0],…,C

[39] ,C

[40] ,…,C

[79] }

[0102] The processor 12 may use the combined pace spectrum data and the minimum autocorrelation value of the combined pace spectrum as gait features to identify the object O. The definition of the combined pace spectrum is described in detail later. The processor 12 may generate M sets of combined Doppler spectrogram data based on M sets of positive velocity energies and M sets of negative velocity energies corresponding to M times in the Doppler spectrogram, and convert N sets of time domain data corresponding to N speeds in the M sets of combined Doppler spectrogram data into N sets of spectrum data respectively. In some embodiments, the processor 12 may divide the energy distribution data of the Doppler spectrogram into positive velocity Doppler spectrum data and negative velocity Doppler spectrum data. The positive velocity Doppler spectrum data and the negative velocity Doppler spectrum data may form a positive velocity Doppler spectrum matrix Dp and a negative velocity Doppler spectrum matrix Dn, respectively, which are expressed by Formula 8 and Formula 9:

[0103]

[0104]

[0105] Each matrix element in the positive velocity Doppler spectrum matrix Dp represents positive velocity energy. The processor 12 can add up all positive velocity energies to generate a total positive velocity energy Esp, which is expressed as Equation 10:

[0106]

[0107] Each matrix element in the negative velocity Doppler spectrum matrix Dn represents negative velocity energy. The processor 12 may add the negative velocity energies to generate a total negative velocity energy Esn, which is expressed as Equation 11:

[0108]

[0109] The processor 12 then divides the element of the Doppler spectrum matrix corresponding to the larger total energy of the positive velocity total energy Esp and the negative velocity total energy Esn by the element at the same position in the Doppler spectrum matrix corresponding to the smaller total energy to generate a combined Doppler spectrum matrix. For example, when the positive velocity total energy is greater than the negative velocity total energy, the Doppler spectrum matrix Dc1 can be expressed as Formula 12:

[0110]

[0111] When the total energy of negative velocity is greater than the total energy of positive velocity, the Doppler spectrum matrix Dc2 can be expressed as formula 13:

[0112]

[0113] In other embodiments, the processor 12 may regard the Doppler spectrum matrix corresponding to the larger total energy of the positive velocity total energy Esp and the negative velocity total energy Esn as a combined Doppler spectrum matrix. The combined Doppler spectrum data may be converted into a combined Doppler spectrum diagram, such as Figure 9 As shown, the horizontal axis represents time and the vertical axis represents the combined energy Ec. The processor 12 can convert the N groups of time domain data corresponding to the N speeds in the combined Doppler spectrum data into N groups of spectrum data according to step S208, and sum up the corresponding energies of the N groups of spectrum data according to step S210 to obtain M combined pace spectrum data C[0] to C[M-1]. For example, the processor 12 can convert the four groups of time domain data corresponding to the four speeds in the combined Doppler spectrum data into four groups of spectrum data, and sum up the energy of the four groups of spectrum data in the frequency domain to obtain 40 combined pace spectrum data C[0] to C

[39] .

[0114] The processor 12 may use Equation 14 to perform an autocorrelation function on the combined paced spectrum data C to generate a plurality of autocorrelation values AV(r). The processor 12 may then calculate the differences between all adjacent autocorrelation values in the plurality of autocorrelation values to generate a plurality of difference values. The processor 12 may use the minimum difference among the plurality of difference values as the minimum autocorrelation difference value.

[0115]

[0116] r is the displacement of the autocorrelation value. For example, if Figure 7 To combine the one-dimensional step spectrum, the processor 12 may use Figure 7 The combined paced spectrum data C in generates a complex number of autocorrelation values AC(r) according to formula 14. Figure 10 The autocorrelation values (0) to AC (49) generated based on the combined step spectrum data C are shown, wherein the horizontal axis represents the displacement r, and the vertical axis represents the adjacent autocorrelation values AC (r). The interval between the adjacent displacements (r+1) and r can be 0.4 Hz, as shown in FIG. Figure 10 As shown, the autocorrelation value AC(r) has peak values at 0Hz, 2Hz, and 4Hz, corresponding to Figure 7 The peak values of the combined step spectrum data C at the resonant frequencies of 0 Hz, 2 Hz, and 4 Hz at the step frequency fc are calculated. That is, the autocorrelation value AC(r) increases at the resonant frequency at the step frequency fc. The processor 12 then calculates 49 differences between two adjacent points AC(r+1) and AC(r) among the autocorrelation values AC(0) to AC(49), and uses the minimum of the 49 differences as the minimum autocorrelation difference. Figure 11 The autocorrelation difference of the combined step spectrum data C is shown, where the horizontal axis represents the displacement r and the vertical axis represents the difference between adjacent autocorrelation values AC(r+1)-AC(r). The difference AC(r+1)-AC(r) has a small difference when the displacement r is 0Hz, 2Hz, and 4Hz, corresponding to Figure 10 The peak values of the autocorrelation values AC(r) at the mid-range displacement r at the resonant frequencies of 0 Hz, 2 Hz, and 4 Hz. The difference between adjacent autocorrelation values, AC(r+1)-AC(r), highlights the strength of each resonant frequency. Stronger resonant frequencies have smaller differences, AC(r+1)-AC(r). Figure 11 The minimum autocorrelation difference is approximately -0.2. The processor 12 may input the combined pace spectrum data and the minimum autocorrelation value into the classifier 120 to classify the object O based on the combined pace spectrum data and the minimum autocorrelation value. Different objects O may have different minimum autocorrelation difference values. For example, the minimum autocorrelation difference value for a person may be approximately -0.2, while the minimum autocorrelation difference value for a dog may be approximately -0.05.

[0117] The processor 12 may use the speed-normalized gait spectrum data as gait features to identify the object O. Using the speed-normalized gait spectrum data can remove or reduce gait features caused by different speeds of the same object, for example, removing or reducing gait features caused by different walking and running speeds of a person. First, the processor 12 may identify the time window of the trunk velocity Vt corresponding to the time index mt and the time window of the adjacent maximum velocity Vm corresponding to the time index mm from the Doppler spectrum data. The processor 12 may also calculate the difference between the time index mt and the time index mm as the speed normalization interval d. For example, if the time index mt is 11 and the time index mm is 1, the speed normalization interval d may be derived as d=11-1=10. The processor 12 may divide each matrix element s[m,h] of the Doppler spectrum matrix Dm in Formula 4 by the corresponding matrix element s[m+d,h] after the speed normalization interval d to generate a speed normalization matrix Dm_vn1, which is expressed as Formula 15:

[0118]

[0119] In some embodiments, the processor 12 may divide each matrix element s[m,h] of the Doppler spectrum matrix Dm in Formula 4 by the corresponding matrix element s[md,h] before the velocity normalization interval d to generate a velocity normalization matrix Dm_vn2, as expressed in Formula 16:

[0120]

[0121] In addition, the processor 12 can scale the Doppler spectrum data and the time axis in equal proportions, as shown in Formula 17:

[0122]

[0123] Where D is the number of time index values after equal scaling within the sampling interval;

[0124] Vt is the trunk velocity;

[0125] Vm is the maximum speed;

[0126] d is the number of time index values in the sampling interval; and

[0127] T is the total time in the sampling interval.

[0128] Figure 12 The figure shows the Doppler spectrum of velocity normalization in the embodiment of the present invention, wherein the horizontal axis represents time t and the vertical axis represents the normalized velocity vn. Figure 12 The energy corresponding to the trunk velocity (2 m / s) in the Doppler spectrum is almost completely removed. The processor 12 can execute steps S208 and S210 based on the velocity-normalized Doppler spectrum data to generate velocity-normalized one-dimensional gait spectrum data. The processor 12 can use the velocity-normalized gait spectrum data as gait features and input them to the classifier 120 to classify the object O.

[0129] The object recognition device 1 and the object recognition method 200 use radar to receive echo signals to generate gait spectrum data, and use the ratio of the secondary part energy to the main part energy in the gait spectrum data, the object movement amplitude, the one-dimensional gait spectrum data, the combined gait spectrum data and the minimum autocorrelation value, or the speed-normalized gait spectrum data as gait features to increase the accuracy and recognition speed of identifying moving objects.

[0130] The above descriptions are merely preferred embodiments of the present invention. Any equivalent changes and modifications made within the scope of the claims of the present invention should fall within the scope of the present invention.

Claims

1. A method for object recognition, characterized in that: Include: generating Doppler spectrum data according to an echo signal, wherein the echo signal is related to the object; Converting N groups of time domain data corresponding to N speeds in the Doppler spectrum data into N groups of spectrum data, where N is a positive integer; combining the N sets of spectrum data to obtain one-step spectrum data; and obtaining a step feature from the step spectrum data to identify the object; The object includes a main part and a secondary part, and obtaining the gait feature from the gait spectrum data to identify the object includes: identifying a fundamental frequency and a main speed of the main part from the cadence spectrum data; generating a movement amplitude according to the fundamental frequency and the main speed; identifying a minor component energy corresponding to the minor component and a major component energy corresponding to the major component from the stepped spectrum data; Producing a proportional relationship between the energy of the secondary portion and the energy of the primary portion; and The object is identified according to the movement amplitude and the proportional relationship.

2. The object recognition method according to claim 1, wherein: Generating the Doppler spectrum data according to the echo signal includes: performing pre-processing on the echo signal to obtain a pre-processed signal; and The pre-processed signal is converted into the Doppler spectrum data.

3. The object recognition method according to claim 2, wherein: Converting the pre-processed signal into the Doppler spectrum data comprises: Performing short-time Fourier transform on the pre-processed signal to generate the Doppler spectrum data.

4. The object recognition method according to claim 2, wherein: Converting the pre-processed signal into the Doppler spectrum data comprises: The pre-processed signal is subjected to wavelet transformation to generate the Doppler spectrum data.

5. The object recognition method according to claim 1, wherein: The step of obtaining the gait feature from the gait spectrum data to identify the object includes: identifying a fundamental frequency and a harmonic frequency from the paced spectrum data; and The object is identified according to the fundamental frequency and the harmonic frequency.

6. The object recognition method according to claim 1, wherein: The N speeds include a plurality of positive speeds and a plurality of negative speeds.

7. The object recognition method according to claim 1, wherein: The converting of the N groups of time domain data corresponding to the N speeds in the Doppler spectrum data into the N groups of spectrum data comprises: generating M sets of combined Doppler spectrum data according to M sets of positive velocity energies and M sets of negative velocity energies corresponding to M times on the Doppler spectrum data, where M is a positive integer; and The N groups of time domain data corresponding to the N speeds in the M groups of combined Doppler spectrum data are respectively converted into N groups of spectrum data.

8. The object recognition method according to claim 7, wherein: The M groups of combined Doppler spectrum data are generated based on the M groups of positive velocity energies and the M groups of negative velocity energies corresponding to the M times on the Doppler spectrum data. Adding the M groups of positive velocity energies to generate a total positive velocity energy; Adding the M groups of negative velocity energies to generate a total negative velocity energy; and An m-th group of velocity energies among the M groups of velocity energies corresponding to a larger total energy of the positive velocity total energy and the negative velocity total energy is divided by an m-th group of velocity energies among the M groups of velocity energies corresponding to a smaller total energy of the positive velocity total energy and the negative velocity total energy to generate an m-th group of combined Doppler spectrum data, where m is an integer and 0≤m <M。 9. The object recognition method according to claim 7, wherein: The M groups of combined Doppler spectrum data are generated based on the M groups of positive velocity energies and the M groups of negative velocity energies corresponding to the M times on the pace spectrum data. The generated M groups of combined Doppler spectrum data include: Adding the M groups of positive velocity energies to generate a total positive velocity energy; Adding the M groups of negative velocity energies to generate a total negative velocity energy; and The M sets of combined Doppler spectrum data are generated according to the M sets of velocity energies corresponding to a larger total energy of the positive velocity total energy and the negative velocity total energy.

10. The object recognition method according to claim 1, wherein: The step of obtaining the gait feature from the gait spectrum data to identify the object includes: performing an autocorrelation process on the paced spectrum data to generate a plurality of autocorrelation values; and The object is identified according to a minimum difference among a plurality of differences between two adjacent points in the plurality of autocorrelation values.

11. The object recognition method according to claim 1, wherein: The method further comprises performing a velocity normalization process on the Doppler spectrum data according to a maximum velocity on the Doppler spectrum data to generate velocity normalized Doppler spectrum data. The step of converting the time domain data corresponding to the N speeds on the Doppler spectrum data into the N groups of spectrum data comprises: The N groups of time domain data corresponding to the N velocities on the velocity-normalized Doppler spectrum data are respectively converted into the N groups of spectrum data.

12. The object recognition method according to claim 1, wherein: Combining the N sets of spectrum data to obtain the pace spectrum data includes: normalizing the N sets of spectrum data; and The N sets of normalized spectrum data are combined to obtain the paced spectrum data.

13. An object recognition device, characterized in that: Include: a radar for receiving an echo signal associated with an object; and a processor coupled to the radar and configured to generate Doppler spectrum data according to an echo signal; Convert the N sets of time domain data corresponding to N speeds in the Doppler spectrum data into N sets of spectrum data respectively, combine the N sets of spectrum data to obtain a step spectrum data, and obtain a step feature from the step spectrum data to identify the object, where N is a positive integer The object comprises a main part and a secondary part, and the processor is further configured to: identifying a fundamental frequency and a main speed of the main part from the cadence spectrum data; generating a movement amplitude according to the fundamental frequency and the main speed; identifying a minor component energy corresponding to the minor component and a major component energy corresponding to the major component from the stepped spectrum data; produces a proportional relationship between the energy of the secondary portion and the energy of the primary portion; and The object is identified according to the movement amplitude and the proportional relationship.

14. The object recognition device according to claim 13, wherein: The slave processor is further configured to: identify a fundamental frequency and a harmonic frequency from the pace spectrum data; and The object is identified according to the fundamental frequency and the harmonic frequency.

15. The object recognition device according to claim 13, wherein: The processor is further configured to generate M sets of combined Doppler spectrum data according to M sets of positive velocity energies and M sets of negative velocity energies corresponding to M times on the Doppler spectrum data, where M is a positive integer; and The N groups of time domain data corresponding to the N speeds in the M combined Doppler spectrum data are respectively converted into N groups of spectrum data.

16. The object recognition device according to claim 13, wherein: The processor is further configured to: perform an autocorrelation process on the paced spectrum data to generate a plurality of autocorrelation values; and The object is identified according to a minimum difference among a plurality of differences between two adjacent points in the plurality of autocorrelation values.

17. The object recognition device according to claim 13, wherein: The processor is further configured to: perform a velocity normalization process on the Doppler spectrum data according to a maximum velocity on the Doppler spectrum data to generate velocity normalized Doppler spectrum data; and The N groups of time domain data corresponding to the N velocities on the velocity-normalized Doppler spectrum data are respectively converted into the N groups of spectrum data.

18. The object recognition device according to claim 13, wherein: The processor is further used to: normalizing the N sets of spectrum data; and The N sets of normalized spectrum data are combined to obtain the paced spectrum data.

Images