Object detection method using radar, object detection system and method for updating the primary classifier in an object detection system

Description

Field of the Invention

[0001] The invention relates to object recognition, and specifically, to an object recognition method using radar, an object recognition system, and a method of updating a primary classifier in an object recognition system.Background of the Invention

[0002] Object recognition is used to identify or classify various moving or stationary target objects such as people, animals, and vehicles in a variety of environments, and is widely used in the fields of autonomous driving, smart home, and robot applications. Related art generally employs image recognition to identify objects. However, clear images are essential for image recognition to ensure the accuracy of recognition. Consequently, the image recognition is unreliable in dark, low-light or complicated light environments. The use of radar can eliminate the interference of light. The radar can be used to detect objects regardless of day or night, in dense woods or tunnel exits where the light changes abruptly. Nevertheless, the accuracy of employing radar to recognize object remains low.

[0003] Document D1(US 2019 / 279366 A1) discloses an artificial neural network for use in a radar, Document D2(Otero M:"Application of a continuous wave radar for human gait recognition", Proceedings of SPIE, IEEE, US, vol. 5809, no. 1, 1 May 2005 (2005-05-01), pages 538-548, XP002384571) discloses a radar for sorting out human gait using Doppler feature, and Document D3(Andric Milenko et al "Analysis of Radar Doppler Signature from Human Data", Radio engineering, vol. 23, no. 1, 1 April 2014 (2014-04-01), pages 11-19, XP055772932) discloses detection of human motion signature in Doppler signal. However, the motion detections in Documents D1 to D3 are insufficiently accurate.

[0004] Therefore, it is desired to develop an object recognition method and an object recognition system to recognize objects using radar, delivering accurate object recognition results without being affected by the light.Summary of the Invention

[0005] The invention is set out in the appended set of claims.Brief Description of the Drawings

[0006] In the following, the invention is further illustrated by way of example, taking reference to the accompanying drawings. Thereof FIG. 1 is a block diagram of an object recognition system according to an embodiment of the invention; FIG. 2 is a schematic diagram of a method of generating training data and training targets and adopted by the object recognition system in FIG. 1; FIG. 3 is a schematic diagram of the primary classifier in FIG. 1; FIG. 4 is a flowchart of an object recognition method according to an embodiment of the invention; and FIG. 5 is a flowchart of a method of updating the primary classifier of the object recognition system in FIG. 1. Detailed Description

[0007] FIG. 1 is a block diagram of an object recognition system 1 according to an embodiment of the invention. The object recognition system 1 utilizes a radar echo signal to recognize a variety of movements of an object, such as a movement of a person, a movement of a dog or a cat, or a movement of a non-living object such as an electric fan. Since various parts of an object move at various velocities and / or various frequencies as the object performs a particular movement, e.g., a person moves his arms and legs at different velocities and frequencies when walking and running, a pet moves its forearms and hind legs at irregular frequencies when moving, a fan rotates at a fixed frequency. Those velocities and / or frequencies are shown in the radar echo signals to serve as features for recognizing or classifying a particular object and / or a particular motion thereof. In practice, the object recognition system 1 trains a classifier in real time according to an echo signal and a training target, using a large quantity of real and representative data sets to reduce a generalization error. The training target includes an object type and / or a motion type. The object type is a person, a cat, a dog, a fan and others, and the motion type is moving forwards, moving backwards, standing still, rotating at a fixed frequency and others. Later, the object recognition system 1 recognizes or classifies an object type and / or a motion type according to echo signals detected by the classifier being trained or completed training.

[0008] The object recognition system 1 includes radar equipment 10, an auxiliary sensor 12, an auxiliary classifier 14, a controller 16 and a memory 18. The radar equipment 10 includes an antenna 100 and a radar transceiver 102. The controller 16 includes a primary classifier 160. The antenna 100 is coupled to the radar transceiver 102. The auxiliary sensor 12 is coupled to the auxiliary classifier 14. The controller 16 is coupled to the radar transceiver 102, the auxiliary classifier 14 and the memory 18.

[0009] The radar equipment 10 is a continuous wave radar, a frequency modulated continuous wave radar, a pulse radar or another type of radar. The antenna 100 transmits a detection signal and receive a corresponding echo signal upon detecting an object. The radar transceiver 102 generates a Doppler signal having a Doppler frequency shift according to the echo signal. The detection signal is a monotone, chirps or periodic pulses. The echo signal is a reflection resulting from the detection signal hitting the object. The Doppler signal is a digital signal or an analog signal. Since a velocity of the object corresponds to the Doppler frequency shift, the controller 16 performs a frequency domain transform on the Doppler signal to generate a frequency domain signal, and determines the velocity of the object according to the frequency domain signal. The frequency domain signal serves as a data set to train the primary classifier 160, and includes Doppler spectrogram data representing variations of velocity components with respect to time and containing patterns to recognize a specific object and / or a specific motion thereof.

[0010] The auxiliary sensor 12 and the auxiliary classifier 14 generate a training target in real time. The auxiliary sensor 12 continuously receives an auxiliary signal, and is a video camera, a camera or a microphone. The type of the auxiliary signal is different from the type of the radar echo signal, for example, the auxiliary signal is a video data stream or an audio data stream. The auxiliary classifier 14 is an image-based or sound-based classification system, e.g., a convolution neural network applicable to computer vision and generating a corresponding object and / or a motion as the training target according to an image data stream, so as to train the primary classifier 160. For example, the training target is a person moving forward or an animal moving. The auxiliary classifier 14 generates a training target having high accuracy, e.g., the accuracy is 95%. The auxiliary classifier 14 and the primary classifier 160 are realized by an artificial neural network such as a multilayer perceptron (MLP) or a convolution neural network, or a classification algorithm such as a support vector machine (SVM). The auxiliary classifier 14 and the primary classifier 160 are stored in the same or different memories, and loaded onto the same or different controllers, microcontrollers, processors, microprocessors or integrated circuits for execution. In some embodiments, the auxiliary classifier 14 and the primary classifier 160 are stored in the memory 18 and loaded onto the controller 16 for execution.

[0011] Since the pattern of the Doppler spectrogram data represents a particular movement performed by a particular object, the primary classifier 160 recognizes the object type and / or the motion type of the object to be classified according to the Doppler spectrogram data. The primary classifier 160 is operated in a training stage or a recognition stage. In the training stage, the antenna 100 transmits a first detection signal and receive a first echo signal upon detecting a first object, the radar transceiver 102 generates a first Doppler signal according to the first echo signal, and the controller 16 generates the first frequency domain signal according to the first Doppler signal. In some embodiments, the controller 16 performs a time-domain-to-frequency-domain transform on the first Doppler signal to generate the first frequency domain signal. Owing to a movement of the first object, the first echo signal has a first Doppler frequency shift relative to the first detection signal. The first frequency signal is a first Doppler frequency domain signal, and carries first Doppler spectrogram data. The first Doppler spectrogram data are energies of velocities with respect to time, and include first data corresponding to M first velocities, where M is a positive integer exceeding 1. The auxiliary sensor 12 receives the auxiliary signal while the antenna 100 receiving the first echo signal, and the auxiliary classifier 14 generates, according to the auxiliary signal, a training target corresponding to the first frequency domain signal. Next, the controller 16 trains the primary classifier 160 by updating at least one parameter of the primary classifier 160 according to the first frequency domain signal and the training target corresponding to the first frequency domain signal. In one embodiment, the controller 16 transforms the first data in the first frequency domain signal and corresponding to M velocities into M sets of first spectrum data, respectively, combines the M sets of first spectrum data to acquire first cadence spectrum data, acquires a first cadence feature from the first cadence spectrum data, and updates the at least one parameter of the primary classifier 160 according to the first cadence feature and the training target of the first frequency domain signal. The first cadence spectrum data include an energy sum over a velocity range corresponding to frequencies in the first Doppler spectrogram data, and include a pattern for use to recognize a specific object, a specific movement and / or a specific movement of the specific object.

[0012] In some embodiments, the antenna 100 receives the first echo signal in a first time interval, the auxiliary sensor 12 receives the auxiliary signal in a second time interval, and the auxiliary classifier 14 generates the training target corresponding to the first frequency domain signal according to the auxiliary signal. The first echo signal is associated with presence of the first object, and the auxiliary signal and the training target are associated with the presence of the first object. In a partially overlapping interval, the first time interval and the second time interval are partially overlapping and the presence of the first object occurs in the partially overlapping interval. The controller 16 determines the presence of the first object according to the first Doppler signal, and the auxiliary classifier 14 inputs the training target corresponding to the first frequency domain signal into the controller 16 according to the presence of the first object. The radar equipment 10 detects the object upon the controller 16 receiving the first Doppler signal, and the primary classifier 160 initiates the training stage. The controller 16 notifies the auxiliary classifier 14 by a control signal to output the training target, and inputs the first frequency domain signal generated by the controller 16 and the training target generated by the auxiliary classifier 14 into the primary classifier 160 to perform training, e.g., updating the at least one parameter of the primary classifier 160. In some embodiments, the controller 16 transmits a trigger signal to the auxiliary classifier 14 to notify the auxiliary classifier 14 to recognize or classify the object type and / or the motion type according the auxiliary signal, and outputs the object type and / or the motion type to the controller 16 as the training target. In other embodiments, upon receiving the first Doppler signal and stopping receiving the first Doppler signal, the controller 16 respectively transmits a segmentation initiation signal and a segmentation termination signal to the auxiliary classifier 14 to notify the auxiliary classifier 14 to segment the auxiliary signal to generate an auxiliary signal segment, to recognize or classify the object type and / or the motion type according the auxiliary signal segment, and to output the object type and / or the motion type to the controller 16 as the training target. The training target is associated with the motion type of the first object, and is of the object type or the motion type.

[0013] In some situations, for example, when the primary classifier 160 completes training or does not operate in the training stage, or the auxiliary sensor 12 or auxiliary classifier 14 is unable / not to be used, the primary classifier 160 enters the recognition stage, and the controller 16 employs the primary classifier 160 to recognize or classify the object type and / or the motion type according to the first frequency domain signal. In the recognition stage, the antenna 100 transmits a second detection signal and receives a second echo signal upon detecting a second object, the radar transceiver 102 generates a second Doppler signal according to the second echo signal, and the controller 16 generates the second frequency domain signal according to the second Doppler signal. In some embodiments, the controller 16 performs a time-domain-to-frequency-domain transform on the second Doppler signal to generate the second frequency domain signal. Owing to a movement of the second object, the second echo signal has a second Doppler frequency shift relative to the second detection signal. The second frequency signal is the second Doppler frequency domain signal, and carries second Doppler spectrogram data. The second Doppler spectrogram data are energies of velocities with respect to time, and include second data corresponding to M second velocities. The controller 16 generates the object classification data corresponding to the second frequency domain signal according to the second frequency domain signal and the at least one parameter of the primary classifier 160, e.g., the parameter updated in the previous training stage.

[0014] In one embodiment, the controller 16 transforms the second data in the second frequency domain signal and corresponding to M velocities into M sets of second spectrum data, respectively, combines the M sets of second spectrum data to acquire second cadence spectrum data, acquires a second cadence feature from the second cadence spectrum data, and generates the object classification data corresponding to the second frequency domain signal according to the at least one parameter of the primary classifier 160 and the second cadence feature. The second cadence spectrum data include an energy sum over a velocity range corresponding to frequencies in the second Doppler spectrogram data, and include a pattern for use to recognize a specific object and / or a specific movement thereof. In some embodiments, the controller 16 further updates the at least one parameter of the primary classifier 160 according to the second frequency domain signal and the object classification data. The object classification data is associated with presence of the second object, and the second object is the same or different from the first object. The object classification data is associated with the motion type of the second object, and is of the object type or the motion type.

[0015] In other embodiments, the controller 16 performs an auto-correlation process on the first cadence spectrum data to generate a plurality of auto-correlations, computes the minimum of differences of two adjacent ones of the plurality of auto-correlations, and employs the minimum as the first cadence feature. Similarly, the controller 16 performs the auto-correlation process on the second cadence spectrum data to generate a plurality of auto-correlations, computes the minimum of differences of two adjacent ones of the plurality of auto-correlations, and employs the minimum as the second cadence feature. The controller 16 updates the at least one parameter of the primary classifier 160 according to the first cadence feature and the training target corresponding to the first frequency domain signal. The controller 16 generates, according to the second cadence feature, the object classification data corresponding to the second frequency domain signal.

[0016] In some embodiments, the controller 16 performs a velocity normalization process on the first spectrogram data according to the maximum velocity of the first spectrogram data to generate first velocity-normalized spectrogram data, and transforms first data of the first velocity-normalized spectrogram data corresponding to the M first velocities into M sets of first spectrum data. Similarly, the controller 16 performs the velocity normalization process on the second spectrogram data according to the maximum velocity of the second spectrogram data to generate second velocity-normalized spectrogram data, and transforms second data of the second velocity-normalized spectrogram data corresponding to the M second velocities into M sets of first spectrum data. In other embodiments, the controller 16 generates first combined spectrogram data according to the positive velocities and negative velocities of the first spectrogram data, and converts first data in the first combined spectrogram data corresponding to M first velocities into M sets of first spectrum data. Likewise, the controller 16 generates second combined spectrogram data according to the positive velocities and negative velocities of the second spectrogram data, and converts second data in the second combined spectrogram data corresponding to M second velocities into M sets of second spectrum data. In one embodiment, the controller 16 computes an energy sum of all the positive velocities and an energy sum of all the negative velocities of the first spectrogram data, determines which of the energy sums is larger, and divides energies of the velocities having the larger energy sum to energies of corresponding velocities having the smaller energy sum to generate the first combined spectrum data. Similarly, the controller 16 computes an energy sum of all the positive velocities and an energy sum of all the negative velocities of the second spectrogram data, determines which of the energy sums is larger, and divides energies of the velocities having the larger energy sum to energies of corresponding velocities having the smaller energy sum, respectively, to generate the second combined spectrum data. In another embodiment, the controller 16 computes an energy sum of all the positive velocities and an energy sum of all the negative velocities of the first spectrogram data, determines which of the energy sums is larger, and generates the first combined spectrum data according to the velocities having the larger energy sum. Similarly, the controller 16 computes an energy sum of all the positive velocities and an energy sum of all the negative velocities of the second spectrogram data, determines which of the energy sums is larger, and generates the second combined spectrum data according to the velocities having the larger energy sum.

[0017] Then, the controller 16 combines the M sets of first spectrum data or the first combined spectrum data to obtain the first cadence spectrum data, and acquires the first cadence feature from the first cadence spectrum data. Likewise, the controller 16 combines the M sets of second spectrum data or the second combined spectrum data to obtain the second cadence spectrum data, and acquires the second cadence feature from the second cadence spectrum data. The controller 16 updates the at least one parameter of the primary classifier 160 according to the first cadence feature and the training target corresponding to the first frequency domain signal. The controller 16 generates the object classification data corresponding to the second frequency domain signal according to the at least one parameter of the primary classifier 160 and the second cadence feature.

[0018] The object recognition system 1 employs the radar equipment 10 and the primary classifier 160 to recognize the object type and / or the motion type without or hardly being affected by the ambient light while providing more accurate object recognition results.

[0019] FIG. 2 is a schematic diagram of a method of generating training data sets and training targets and adopted by the object recognition system 1. In the embodiment, the radar equipment 10 is a continuous wave radar, the auxiliary sensor 12 is a video camera, and the auxiliary classifier 14 is a convolution neural network configured to classify objects according to images. In a time interval T, the auxiliary sensor 12 receives and records a video data stream in real time, while the radar equipment 10 continuously detecting an object. Upon the radar equipment 10 detecting an object for the first time, the controller 16 notifies the auxiliary classifier 14 to segment the video data stream to generate a video data segment S1 and generates a training target T1 (e.g., a person moving forward) according to the video data segment S1, the controller 16 generates a data set D1 (e.g., Doppler spectrogram data corresponding to the person moving forward), and then the controller 16 stores the training target T1 and the data set D1 as a training file F1 for use by the primary classifier 160 in subsequence training, or directly outputs the training target T1 and the data set D1 to the primary classifier 160 to perform real-time training. Upon the radar equipment 10 detecting an object for the second time, the controller 16 notifies the auxiliary classifier 14 to segment the video data stream to generate a video data segment S2 and generates a training target T2 (e.g., a person standing still) according to the video data segment S2, the controller 16 generates a data set D2 (e.g., Doppler spectrogram data corresponding to the person standing still), and then the controller 16 stores the training target T2 and the data set D2 as a training file F2 for use by the primary classifier 160 in subsequence training, or directly outputs the training target T2 and the data set D2 to the primary classifier 160 to perform real-time training. Upon the radar equipment 10 detecting an object for the third time, the controller 16 notifies the auxiliary classifier 14 to segment the video data stream to generate a video data segment S3 and generates a training target T3 (e.g., a pet moving) according to the video data segment S3, the controller 16 generates a data set D3 (e.g., Doppler spectrogram data corresponding to the pet moving), and then the controller 16 stores the training target T3 and the data set D3 as a training file F3 for use by the primary classifier 160 in subsequence training, or directly outputs the training target T3 and the data set D3 to the primary classifier 160 to perform real-time training. Upon the radar equipment 10 detecting an object for the fourth time, the controller 16 notifies the auxiliary classifier 14 to segment the video data stream to generate a video data segment S4 and generates a training target T4 (e.g., a person moving backward) according to the video data segment S4, the controller 16 generates a data set D4 (e.g., Doppler spectrogram data corresponding to the person moving backward), and then the controller 16 stores the training target T4 and the data set D4 as a training file F4 for use by the primary classifier 160 in subsequence training, or directly outputs the training target T4 and the data set D4 to the primary classifier 160 to perform real-time training. In another embodiment, upon the auxiliary classifier 14 detecting objects, the controller 16 notifies the auxiliary classifier 14 to segment the video data stream to generate the video data segments S1 to S4. In one embodiment, the length of the video data segments S1 to S4 are determined by one or a combination of the following: a fixed value, a time interval in which the radar equipment 10 detects the presence of the object, and a time interval in which the auxiliary classifier 14 detects the presence of the object.

[0020] FIG. 3 is a schematic diagram of the primary classifier 160 adopting a multilayer perceptron. The primary classifier 160 includes an input layer, a hidden layer and an output layer. In some embodiments, multiple hidden layers are applied. The output layer includes input nodes Nil to Nin, the hidden layer includes hidden node Nh1 to Nhp, and the output layer includes output nodes No1 to Nom, wherein n, p and m are positive integers. The input nodes Nil to Nin receive a training data set including inputs X1 to Xn, respectively. The hidden nodes Nh1 to Nhp receive linear weighted sums of the inputs X1 to Xn via n connection parameters and generate hidden node outputs h1 to hp via activation functions, respectively. For example, the hidden node Nh1 receives a linear weighted sum (X1*Wi1+ X2*Wi2+...+Xn*Win) and generates a hidden node output h1 via an activation function tanh. The output nodes No1 to Nom receive a linear weighted sum of all hidden node outputs via p connection parameters and generate estimated outputs Y1 to Ym via activation functions, respectively. For example, the output node No1 receives the linear weighted sum (h1*Wh1+ h2*Wh2+ h3*Wh3+...+hp*Whp) and generates the estimated output Y1 via the activation function tanh. The controller 16 adjusts or updates the at least one parameter in the primary classifier 160 to reduce errors δ1 to δm between the estimated outputs Y1 to Ym and corresponding training targets T1 to Tm, respectively. The inputs X1 to Xn are the training data set generated by the controller 16, the training targets T1 to Tm are the object type and / or the motion type generated by the auxiliary classifier 14. When using the Doppler spectrogram data as the training data set for training the primary classifier 160, the inputs X1 to Xn are n velocity components at n points in time. When using the first cadence spectrum data as the training data set for training the primary classifier 160, the inputs X1 to Xn are n sums over energies of velocities at n points in time. In one embodiment, the inputs X1 to Xn are the aforementioned first cadence features.

[0021] While the embodiment in FIG. 3 adopts the multilayer perceptron as the primary classifier 160, other types of artificial neural networks or other classification algorithms may also be adopted.

[0022] FIG. 4 is an overview flowchart of an object recognition method 4 according to an embodiment of the invention. The object recognition method 4 is applicable to the object recognition system 1, and includes Steps S400 to S408. Steps S400 to S406 are used to train the primary classifier 160, and Step S408 is used to recognize the object type and / or the motion type. Steps S400 to S408 are detailed as below: Step S400:Transmit the first detection signal, receive the first echo signal when the first object is detected in the first time interval, and generate the first frequency domain signal according to the first echo signal; proceed to Step S404;Step S402:Receive the auxiliary signal in the second time interval, and generate the training target corresponding to the first frequency domain signal according to the auxiliary signal, the first time interval being partially overlapping with the second time interval;Step S404:Update the at least one parameter of the primary classifier 160 according to the first frequency domain signal and the training target corresponding to the first frequency domain signal;Step S406:Transmit the second detection signal, receive the second echo signal whenthe second object is detected, and generate the second frequency domain signal according to the second echo signal;Step S408:Generate the object classification data corresponding to the second frequency domain signal according to the second frequency domain signal and the at least one parameter of the primary classifier 160.

[0023] Steps S400 to S408 are explained in the preceding paragraphs, the details thereof are not repeated here. Steps S400 to S408 are used to train the primary classifier 160 to identify objects, and after the training is completed, the radar equipment 10 is used to directly identify the object type and / or the motion type without being affected by ambient light.

[0024] FIG. 5 is an overview flowchart of a method 5 of updating the primary classifier of the object recognition system 1. The method 5 is applicable to the object recognition system 1, and includes Steps S500 to S504 for use to train the primary classifier 160. Steps S500 to S504 are detailed as below: Step S500:Transmit the first detection signal, receive the first echo signal when the first object is detected in the first time interval, and generate the first frequency domain signal according to the first echo signal; proceed to Step S504;Step S502:Receive the auxiliary signal in the second time interval, and generate the training target corresponding to the first frequency domain signal according to the auxiliary signal, the first time interval being at least partially overlapping with the second time interval;Step S504Update at least one parameter of the primary classifier 160 according to the first frequency domain signal and the training target corresponding to the first frequency domain signal.

[0025] In above Steps, the first echo signal is associated with presence of the first object, the auxiliary signal is associated with the presence of the first object, and the presence of the first object occurs during the at least partially overlapping interval. Steps S500 to S504 are explained in the preceding paragraphs, the details thereof will not be repeated here. Where the external environment is appropriate, for example, when the ambient light is sufficient or the ambient noise is low, in Steps S500 to S504, the auxiliary sensor 12 and the auxiliary classifier 14 generate training targets in real time to train primary classifier 160 to recognize objects, thereby performing supervised learning on the object recognition system 1.

[0026] The object recognition method 4 and the object recognition system 1 employ the radar equipment 10 to recognize the object type and / or the motion type without or hardly being affected by the ambient light while providing more accurate object recognition results. The object recognition system 1 and the method 5 of updating the primary classifier 160 in the object recognition system generate training targets in real time to provide supervised learning to the object recognition system 1, enhancing a success rate of recognizing an object.

Claims

1. An object recognition method of an object identification system (1), the object recognition method comprising: receiving by an antenna (100) a first echo signal during a first interval; receiving by an auxiliary sensor (12) an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; generating by a controller (16) a first frequency domain signal according to the first echo signal; generating by an auxiliary classifier (14) a training target (T1, T2, T3, T4) corresponding to the first frequency domain signal according to the auxiliary signal; transforming by the controller (16) first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, wherein M is a positive integer exceeding 1; combining by the controller (16) the M sets of first spectrum data to acquire first cadence spectrum data; acquiring by the controller (16) a first cadence feature from the first cadence spectrum data; updating by the controller (16) at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal; generating by the controller (16) a second frequency domain signal according to a second echo signal; computing by the controller (16) an energy sum of positive velocities and an energy sum of negative velocities for second spectrogram data in the second frequency domain signal; determining by the controller (16) which of the energy sums is larger; dividing by the controller (16) energies of velocities having a larger energy sum to energies of corresponding velocities having a smaller energy sum, respectively, to generate second combined spectrum data; transforming by the controller (16) second data corresponding to M second velocities and in the second combined spectrogram data into M sets of second spectrum data; combining by the controller (16) the M sets of second spectrum data to acquire second cadence spectrum data; acquiring by the controller (16) a second cadence feature from the second cadence spectrum data; and generating by the controller (16) object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160), wherein the object classification data is associated with presence of a second object.

2. An object recognition method of an object identification system (1), the object recognition method comprising: receiving by an antenna (100) a first echo signal during a first interval; receiving by an auxiliary sensor (12) an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; generating by a controller (16) a first frequency domain signal according to the first echo signal; generating by an auxiliary classifier (14) a training target (T1, T2, T3, T4) corresponding to the first frequency domain signal according to the auxiliary signal; transforming by the controller (16) first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, wherein M is a positive integer exceeding 1; combining by the controller (16) the M sets of first spectrum data to acquire first cadence spectrum data; acquiring by the controller (16) a first cadence feature from the first cadence spectrum data; updating by the controller (16) at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal; generating by the controller (16) a second frequency domain signal according to a second echo signal; computing by the controller (16) an energy sum of positive velocities and an energy sum of negative velocities for second spectrogram data in the second frequency domain signal; determining by the controller (16) which of the energy sums is larger; generating by the controller (16) second combined spectrogram data according to velocities having a larger energy sum; transforming by the controller (16) second data corresponding to M second velocities and in the second combined spectrogram data into M sets of second spectrum data; combining by the controller (16) the M sets of second spectrum data to acquire second cadence spectrum data; acquiring by the controller (16) a second cadence feature from the second cadence spectrum data; and generating by the controller (16) object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160), wherein the object classification data is associated with presence of a second object.

3. An object recognition method of an object identification system (1), the object recognition method comprising: receiving by an antenna (100) a first echo signal during a first interval; receiving by an auxiliary sensor (12) an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; generating by a controller (16) a first frequency domain signal according to the first echo signal; generating by an auxiliary classifier (14) a training target (T1, T2, T3, T4) corresponding to the first frequency domain signal according to the auxiliary signal; transforming by the controller (16) first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, wherein M is a positive integer exceeding 1; combining by the controller (16) the M sets of first spectrum data to acquire first cadence spectrum data; acquiring by the controller (16) a first cadence feature from the first cadence spectrum data; updating by the controller (16) at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal; generating by the controller (16) a second frequency domain signal according to a second echo signal; transforming by the controller (16) second data corresponding to M second velocities and in the second frequency domain signal into M sets of second spectrum data; combining by the controller (16) the M sets of second spectrum data to acquire second cadence spectrum data; performing by the controller (16) an autocorrelation process on the second cadence spectrum data to generate a plurality of autocorrelation values; computing by the controller (16) a minimum difference of a plurality of differences between two adjacent ones of the plurality of autocorrelation values to employ the minimum difference as a second cadence feature; and generating by the controller (16) object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160), wherein the object classification data is associated with presence of a second object.

4. The object recognition method of any of claims 1-3 , further comprising updating the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160) according to the second frequency domain signal and the object classification data.

5. The object recognition method of any of claims 1-3, whereby the first frequency domain signal is a first Doppler frequency domain signal, the second frequency domain signal is a second Doppler frequency domain signal, the first echo signal has a first Doppler frequency shift relative to a first detection signal and corresponding to a movement of a first object, the second echo signal has a second Doppler frequency shift relative to a second detection signal and corresponding to a movement of the second object, the training target (T1, T2, T3, T4) is associated with a motion type of the first object, and the object classification data is associated with a motion type of the second object.

6. The object recognition method of any of claims 1-3, further comprising determining presence of a first object according to the first echo signal; and inputting the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal according to the presence of the first object.

7. The object recognition method of any of claims 1-3, whereby a type of the auxiliary signal is different from a type of the first echo signal, and the training target (T1, T2, T3, T4) and the object classification data respectively comprise an object type or a motion type.

8. The object recognition method of any of claims 1-3 , further comprising performing a velocity normalization process on the second spectrogram data according to a maximum velocity of the second spectrogram data to generate second velocity-normalized spectrogram data; wherein transforming the second data corresponding to the M second velocities and in the second frequency domain signal into the M sets of second spectrum data comprises: transforming the second data corresponding to the M second velocities and in the second velocity-normalized spectrogram data into M sets of second spectrum data, respectively.

9. An object identification system (1) comprising an antenna (100) configured to transmit a first detection signal and receive a first echo signal upon detecting a first object in a first interval; an auxiliary sensor (12) configured to receive an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; an auxiliary classifier (14) coupled to the auxiliary sensor (12) and configured to generate a training target (T1, T2, T3, T4) corresponding to a first frequency domain signal according to the auxiliary signal; a controller (16) coupled to the antenna (100) and the auxiliary classifier (14), and configured to generate a frequency domain signal according to the first echo signal, transform first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, combine the M sets of first spectrum data to acquire first cadence spectrum data, acquire a first cadence feature from the first cadence spectrum data, and update at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal, M being a positive integer exceeding 1; wherein the antenna (100) is further configured to transmit a second detection signal and receive a second echo signal upon detecting a second object; the controller (16) is further configured to generate a second frequency domain signal according to the second echo signal, compute an energy sum of positive velocities and an energy sum of negative velocities for second spectrogram data in the second frequency domain signal, determine which of the energy sums is larger, divide energies of velocities having a larger energy sum to energies of corresponding velocities having a smaller energy sum, respectively, to generate second combined spectrum data, transform second data corresponding to M second velocities and in the second combined spectrogram data into M sets of second spectrum data, combine the M sets of second spectrum data to acquire second cadence spectrum data, acquire a second cadence feature from the second cadence spectrum data, and generate object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160); and the object classification data is associated with presence of the second object.

10. An object identification system (1) comprising: an antenna (100) configured to transmit a first detection signal and receive a first echo signal upon detecting a first object in a first interval; an auxiliary sensor (12) configured to receive an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; an auxiliary classifier (14) coupled to the auxiliary sensor (12) and configured to generate a training target (T1, T2, T3, T4) corresponding to a first frequency domain signal according to the auxiliary signal; a controller (16) coupled to the antenna (100) and the auxiliary classifier (14), and configured to generate a frequency domain signal according to the first echo signal, transform first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, combine the M sets of first spectrum data to acquire first cadence spectrum data, acquire a first cadence feature from the first cadence spectrum data, and update at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal, M being a positive integer exceeding 1; wherein the antenna (100) is further configured to transmit a second detection signal and receive a second echo signal upon detecting a second object; the controller (16) is further configured to generate a second frequency domain signal according to the second echo signal, compute an energy sum of positive velocities and an energy sum of negative velocities for second spectrogram data in the second frequency domain signal, determine which of the energy sums is larger, generate second combined spectrogram data according to velocities having a larger energy sum, transform second data corresponding to M second velocities and in the second combined spectrogram data into M sets of second spectrum data, combine the M sets of second spectrum data to acquire second cadence spectrum data, acquire a second cadence feature from the second cadence spectrum data, and generate object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160); and the object classification data is associated with presence of the second object.

11. An object identification system (1) comprising: an antenna (100) configured to transmit a first detection signal and receive a first echo signal upon detecting a first object in a first interval; an auxiliary sensor (12) configured to receive an auxiliary signal during a second interval, wherein the first interval and the second interval are at least partially overlapping; an auxiliary classifier (14) coupled to the auxiliary sensor (12) and configured to generate a training target (T1, T2, T3, T4) corresponding to a first frequency domain signal according to the auxiliary signal; a controller (16) coupled to the antenna (100) and the auxiliary classifier (14), and configured to generate a frequency domain signal according to the first echo signal, transform first data corresponding to M first velocities and in the first frequency domain signal into M sets of first spectrum data, combine the M sets of first spectrum data to acquire first cadence spectrum data, acquire a first cadence feature from the first cadence spectrum data, and update at least one parameter (Wi1 to Win, Wh1 to Whp) of a primary classifier (160) according to the first cadence feature and the training target (T1, T2, T3, T4) corresponding to the first frequency domain signal, M being a positive integer exceeding 1; wherein the antenna (100) is further configured to transmit a second detection signal and receive a second echo signal upon detecting a second object; the controller (16) is further configured to generate a second frequency domain signal according to the second echo signal, transform second data corresponding to M second velocities and in the second frequency domain signal into M sets of second spectrum data, combine the M sets of second spectrum data to acquire second cadence spectrum data, perform an autocorrelation process on the second cadence spectrum data to generate a plurality of autocorrelation values, compute a minimum difference of a plurality of differences between two adjacent ones of the plurality of autocorrelation values to employ the minimum difference as a second cadence feature; and generate object classification data corresponding to the second frequency domain signal according to the second cadence feature and the at least one parameter (Wi1 to Win, Wh1 to Whp) of the primary classifier (160); and the object classification data is associated with presence of the second object.