Target classification method, device and storage medium

Abstract

The present invention discloses a target classification method, device, and storage medium. For each target, the method aims to input the observation data sequence corresponding to the target into a hidden Markov model, and then use the hidden Markov model to decode the tracking trajectory data of the radar detection target to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states. For each state of the first target probability function, a logarithm operation is performed on the first target probability function to obtain a second target probability function, and a classifier function is constructed based on a portion of the second target probability function. The classifier function is then determined based on the state value corresponding to the minimum value output by the classifier function. The technical solution of the embodiments of the present invention significantly reduces the computational overhead of using a hidden Markov model to classify radar trajectory data.

Description

Technical Field

[0001] The present application belongs to the field of radar data processing technology, and specifically relates to a target classification method, device and storage medium. Background Art

[0002] Roadside millimeter-wave radar, operating in the 79-81 GHz frequency range, is primarily used to observe road targets such as vehicles, non-motorized vehicles, and pedestrians. Millimeter-wave radar excels in range, angle, and velocity measurement, and the trajectory data generated by radar tracking is a time series pattern (a time series pattern is one in which the spatial distribution of a pattern's features is time-series).

[0003] The Hidden Markov Model is a time series model that excels at processing time series data, meaning each time node has a data set. Therefore, it can be applied to speech signal recognition and classification models. However, the decoding calculations used to classify radar tracking trajectory data using this model are very complex, increasing the computational overhead of the model. Summary of the Invention

[0004] The purpose of this application is to provide a target classification method, device and storage medium that can solve the above technical problems.

[0005] According to one aspect of the present invention, a method for object classification is provided, the method comprising:

[0006] Acquiring tracking trajectory data of a target detected by the radar and generating a tracking data sequence, and extracting at least one feature vector corresponding to the target and used to characterize the target from the tracking data sequence to construct an observation data sequence;

[0007] For each target, inputting the observation data sequence corresponding to the target into a hidden Markov model, and using the hidden Markov model to decode the tracking trajectory data of the target detected by the radar to obtain a first target probability function of the value of the observation data sequence corresponding to all different initial states;

[0008] For each state, performing a logarithm operation on the first target probability function to obtain a second target probability function, and constructing a classifier function based on a portion of the second target probability function;

[0009] The category of the current tracking trajectory data of the radar is determined according to the state value corresponding to the minimum value output by the classifier function.

[0010] Furthermore, for each target, the observation data sequence corresponding to the target is input into a hidden Markov model, and the tracking trajectory data of the target detected by the radar is decoded using the hidden Markov model to obtain the first target probability function of the value of the observation data sequence corresponding to all different initial states, including:

[0011] Let the initial state , , represents the number of states of the hidden Markov chain, and the observation data sequence is expressed as hour,

[0012] in, is a set of feature vectors, Expressed as [RCS ,xSize ,ySize ], , Indicates the frame number;

[0013] Set the transition probability matrix in the model to the identity matrix to calculate the hidden state values for all target classification states.

[0014] Furthermore, the method of decoding the tracking trajectory data of the target detected by the radar to obtain the hidden state sequence with the highest probability in the current radar trajectory data, and calculating the probabilities of the values of the observation data sequence corresponding to all different initial states includes:

[0015] The first target probability function of the value of the observed data sequence corresponding to all different initial states is calculated according to the following formula:

[0016] ;

[0017] ;

[0018] ;

[0019] in, represents the first target probability function of each state, , represents the number of states of the hidden Markov chain, represents the conditional probability, represents the observation data sequence, It indicates the probability of the observed data sequence X appearing when the model is in the hidden state H. is the initial state probability of the model, represents the probability density function, M represents the number of observations corresponding to each state, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0020] For each state, the method of performing a logarithmic operation on the first target probability function to obtain a second target probability function includes:

[0021] The first step is to , , represents the number of states of the hidden Markov chain, the second target probability function is obtained by the following formula:

[0022] ;

[0023] The second step is to process the second target probability function to obtain:

[0024] ;

[0025] The third step is to further process the second target probability function to obtain:

[0026] ;

[0027] Step 4: ,in,

[0028] ,in, is a constant term, ;

[0029] in, represents the second target probability function of each state, , represents the number of states of the hidden Markov chain, is the initial state probability of the model, represents the number of frames, M represents the number of observations corresponding to each state, represents the observation data sequence, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0030] Furthermore, constructing a classifier function based on a portion of the second target probability function includes:

[0031] make , where the classifier function is obtained by the following formula:

[0032] ;

[0033] in, Represents the classifier function for each state, is the initial state probability of the model, , represents the number of states of the hidden Markov chain, Indicates the number of frames, is the mean vector, M represents the number of observations corresponding to each state, , is the covariance matrix, yes The inverse matrix of the covariance matrix is defined as .

[0034] Furthermore, the method further comprises:

[0035] subtracting a current frame of the classifier function from a previous frame of the classifier function to construct a value-added function between two adjacent frames of the classifier function;

[0036] Calculating the values of the classifier function in the previous frame for all different initial states and caching them as reference values for the current frame;

[0037] The reference value and the value calculated by the value-added function are added to output the value of the classifier function in real time.

[0038] Furthermore, the method further comprises:

[0039] The category of the current tracking trajectory data of the radar is determined according to the state value corresponding to the minimum value of the current frame output by the classifier function.

[0040] Furthermore, the method further comprises:

[0041] The values of the current frame of the classifier function in all different initial states are cached and updated as the reference value of the next frame of the classifier function.

[0042] Furthermore, the method of subtracting the current frame of the classifier function from the previous frame of the classifier function to construct a value-added function between two adjacent frames of the classifier function includes:

[0043] The previous frame of the classifier function is defined as , the previous frame of the classifier function is represented by the following formula:

[0044] ;

[0045] The current frame of the classifier function is defined as , the current frame of the classifier function is represented by:

[0046] ;

[0047] right Processing yields:

[0048] ;

[0049] This is converted to:

[0050] ;

[0051] The value-added function is defined as ,but ,

[0052] ;

[0053] in, represents the value-added function, , represents the number of states of the hidden Markov chain, is the initial state probability of the model, Indicates the number of frames, represents the observation data sequence, represents the observation data sequence of the first frame, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0054] According to another aspect of the present invention, there is provided an object classification device, the device comprising:

[0055] an acquisition module, configured to acquire tracking trajectory data of a target detected by the radar and generate a tracking data sequence, and extract at least one feature vector corresponding to the target from the tracking data sequence to construct an observation data sequence;

[0056] a decoding module for inputting, for each target, an observation data sequence corresponding to the target into a hidden Markov model, and decoding the tracking trajectory data of the target detected by the radar using the hidden Markov model to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states;

[0057] a classifier function construction module, configured to perform a logarithmic operation on a first target probability function for each state to obtain a second target probability function, and construct a classifier function based on a portion of the second target probability function;

[0058] The determination module is used to determine the category of the current tracking trajectory data of the radar according to the state value corresponding to the minimum value output by the classifier function.

[0059] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the target classification method is implemented.

[0060] The beneficial effects of the present invention are as follows: compared with the existing technology, the target classification model training method and target classification method, device and storage medium provided by the present invention solve the problem that radar in current technology has difficulty in classifying detected targets and the problem of low target classification accuracy. By utilizing the tracking data sequence of each target within the radar monitoring range for classification, all information captured by the radar during the target's historical process can be fully utilized, thereby significantly improving the accuracy of radar target classification.

[0061] Furthermore, the Hidden Markov Model (HMM) was simplified, reducing the computational complexity of model decoding. A continuous HMM approach was used to adapt to the characteristics of millimeter-wave radar data and to classify radar target trajectory data. Using the HMM, the classification of traffic millimeter-wave radar data was very effective, with a classification recall rate exceeding 95%.

[0062] Furthermore, the embodiments of the present invention significantly reduce the computational overhead of classifying radar trajectory data using a hidden Markov model.

[0063] Furthermore, the embodiment of the present invention can calculate and update the target category information in real time during the process of radar tracking the target trajectory, without having to wait until the target tracking trajectory is completed before classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0065] Figure 1 This is a flow chart of a target classification model training method provided by an embodiment of the present invention.

[0066] Figure 2 This is a flow chart of a target classification method provided by an embodiment of the present invention.

[0067] Figure 3 This is a structural block diagram of a target classification device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following, in conjunction with the accompanying drawings, provides a clear and complete description of the technical solutions in the embodiments of the present application. It should be noted that the described embodiments are only a portion of the embodiments of the present application, and are not intended to be exhaustive. All other embodiments derived by persons skilled in the art based on the embodiments of the present application without inventive effort are intended to fall within the scope of protection of the present application. Furthermore, in the description of the present application, the term "including" means "including but not limited to." The terms "first," "second," and "third," etc., are used merely as designations and do not impose numerical requirements or establish a sequence. Various embodiments of the present application may be presented in the form of a range. It should be understood that describing in a range format is merely for convenience and brevity and should not be construed as a rigid limitation on the scope of the present application. Therefore, the range description should be considered to specifically disclose all possible subranges and individual numerical values within the range. For example, a range description of 1 to 6 should be considered to specifically disclose subranges, such as 1 to 3, 1 to 4, 1 to 5, 2 to 4, 2 to 6, 3 to 6, etc., as well as individual numerical values within the range, such as 1, 2, 3, 4, 5, and 6, regardless of the range. Additionally, whenever a numerical range is indicated herein, it is meant to include any cited numeral (fractional or integral) within the indicated range.

[0069] Figure 1 This is a flow chart of a target classification model training method provided by an embodiment of the present invention.

[0070] refer to Figure 1 As shown, an embodiment of the present invention provides a target classification model training method comprising the following steps:

[0071] Step S11, obtaining tracking trajectory data of a target detected by the radar and generating a tracking data sequence, and extracting at least one feature vector corresponding to the target and used to characterize the target from the tracking data sequence to construct an observation data sequence;

[0072] Step S12: performing target classification processing on the observed data sequence based on preset rules, and calculating the actual occurrence probability of each category, and using the actual occurrence probability of each category as the initial state probability of each state;

[0073] Step S13, estimating the parameters of the hidden Markov model based on the initial state probabilities of each state: ,in, is the initial state probability of the model corresponding to each category at the initial moment, A is the transition probability matrix, and B is the observation probability matrix.

[0074] Steps S11 to S13 will be described in detail below.

[0075] In step S11 , the tracking trajectory data of the target detected by the radar is obtained and a tracking data sequence is generated.

[0076] Traffic millimeter-wave radars generally sample at fixed time intervals, such as 100ms. This means that every 100ms, the radar will scan the current monitoring area, generate point cloud data for the current monitoring area, and then generate tracking trajectory data for the detected target based on the point cloud data and historical data. It should be noted that the tracking trajectory data is composed of tracking point data. Each tracking point data represents a mobile target in the radar monitoring area and is marked with a unique ID tag. The collection of tracking point data for all mobile targets in the monitoring area at the current time is the tracking trajectory data for the current frame. After 100ms, the radar will continue to generate a frame of tracking point data. If the ID tag of each mobile target in the current frame is consistent with the ID tag of its previous frame, the movement trajectory data of each mobile target within the radar monitoring range can be obtained.

[0077] Although each tracked target can be classified based on each frame of tracking data acquired, due to the high noise in the data, the classification accuracy will be very low if the target is classified based on only one frame of tracking data. To address the problem of low radar target classification accuracy, in an embodiment of the present invention, the millimeter-wave radar generates a frame of tracking data after each sampling, and as a fixed time passes, it can generate tracking trajectory data corresponding to several frames. Within this time period, all tracking points of the same target are arranged in chronological order to obtain the tracking data sequence of the current target within this time period. By using the tracking data sequence of each target within the radar monitoring range for classification, all information captured by the radar during the historical movement of the target can be fully utilized, thereby significantly improving the accuracy of radar target classification.

[0078] At least one feature vector corresponding to and used to characterize the target is extracted from the tracking data sequence to construct an observation data sequence.

[0079] Since the tracking data sequence of each target is generated by the entire trajectory of the radar tracking target, this tracking data sequence can be used for target classification. The tracking data sequence is the radar tracking trajectory data r1 starting from the time t1 when the target is detected by the radar to the last frame time t before the target is lost from the radar monitoring range. T Radar tracking trajectory data r T , a total of T frames of data. That is, the trajectory of the current target is available To express.

[0080] Radar tracking data Contains a series of radar-generated features, such as the position information x, y of the corresponding target, and the velocity information vx, vy of the corresponding target. Therefore, a portion of feature vectors suitable for target classification is screened from the radar tracking trajectory data to construct an observation data sequence; for example, in this embodiment of the present invention, the three features of RCS (Radar Cross Section) value, xSize value and ySize value are used as the observation state value M (the number of possible observation values corresponding to each state, that is, in this embodiment of the present invention, ), so after feature screening, the trajectory R of the current target is used to generate the observation data sequence ,in, is a set of feature vectors, Expressed as [RCS ,xSize ,ySize ], , Indicates the frame number.

[0081] It should be understood that the RCS value is primarily related to the strength of the target's scattered echo signal and the target's scattering characteristics. Factors influencing the RCS value include the target's structure and surface medium, radar frequency, polarization method, and radar line of sight (target attitude angle and incident direction of the electromagnetic wave). Radar tracking also provides the target's xSize and ySize values. These values represent the estimated size of the target along the X and Y axes of the radar coordinate system, respectively. These values are output by the radar algorithm and are not detailed here.

[0082] In step S12, target classification processing is performed on the observation data sequence based on preset rules, and the actual occurrence probability of each category is counted, and the actual occurrence probability of each category is used as the initial state probability of each state.

[0083] For example, in an embodiment of the present invention, the tracking trajectory data of the target detected by the radar is obtained, the tracking trajectory data of multiple targets are collected and the tracking trajectory data are classified and labeled using a manual labeling method; wherein the manually labeled category information includes large vehicles, small vehicles, pedestrians, and non-motor vehicles. Thus, the number of states of the hidden Markov chain is N=4, and the initial state probability of each state is , in this embodiment, It is equivalent to the probability of each category actually appearing in the model at the initial moment. Specific calculation method: collect a large amount of tracking trajectory data of the radar detection targets, manually mark the categories of the radar tracking trajectory data, and count the probability of each category appearing in the tracking trajectory data as the initial state probability of the model corresponding to each category at the initial moment. That is to say, based on the current traffic scenario, by collecting a large amount of tracking trajectory data of the radar detection targets, the initial state probabilities of large vehicles, small vehicles, pedestrians and non-motor vehicles that actually appear can be classified and counted. For example, the initial state probabilities of large vehicles, small vehicles, pedestrians and non-motor vehicles are .

[0084] In step S13, the parameters of the Hidden Markov Model (HMM) are estimated based on the initial state probabilities of each state: ,in, is the initial state probability of the model corresponding to each category at the initial moment, A is the transition probability matrix, and B is the observation probability matrix; and the transition probability matrix is set to the unit matrix.

[0085] In a conventional HMM chain, A represents the state change between hidden states. In this embodiment, considering that the category of the target tracked by the millimeter-wave radar is generally fixed, that is, the category of the same target does not jump, for the same target, the initial hidden state is set to one of the four states mentioned above: large vehicle, small car, pedestrian, and non-motor vehicle. Therefore, the transition probability matrix under the current model can be simplified to the identity matrix, that is, the transition probability matrix is expressed as follows:

[0086] .

[0087] In the embodiment of the present invention, exemplarily, when the transition probability matrix A is set to the identity matrix, at this time, the hidden state value of each state is equal to the initial state value, that is, the initial state , , represents the number of states of the hidden Markov chain, and the hidden state sequence is Therefore, we only need to calculate the probabilities of N initial state values to complete the estimation of all possible hidden state sequences. For all possible hidden state sequences, we estimate the probability of the current observation data sequence. The hidden state sequence with the highest probability is the hidden state sequence corresponding to the current observation sequence.

[0088] If the transition probability matrix A is a non-unit matrix, that is, a conventional probability distribution, then when the initial state , , When represents the number of states of the hidden Markov chain, then will probably be The specific probability of occurrence of any value in is determined by the probability distribution of the corresponding i-th row in the transition probability matrix A. And so on, until the last hidden state of the hidden state sequence Considering that there are N possibilities for the initial state, the total number of hidden state sequences is possibilities, we need to calculate them separately The probability of the observation data sequence corresponding to the hidden state appears, and the The maximum value among the probabilities. Obviously, the calculation method for model decoding will become very cumbersome at this time, and the resulting computational overhead is extremely huge. Of course, there are some methods for quickly calculating the hidden state sequence corresponding to the maximum probability, but there is still a certain computational overhead. Such fast calculation methods are basic industry knowledge, so they will not be detailed here.

[0089] It should be understood that in the embodiment of the present invention, the unit matrix is used as the transfer probability matrix. Compared with the original decoding, which has a large amount of calculation and some calculation methods are also very complicated, it can simplify the model decoding operation and make the model decoding operation relatively simple.

[0090] In the embodiment of the present invention, since the observation data sequence The observed feature values (RCS value, xSize value, ySize value) involved in are all continuous values, not discrete values, so the discrete model in the original hidden Markov model cannot be used directly. Because each row in the observation probability matrix in the original discrete model All are hidden states The probability distribution of the observed eigenvalues.

[0091] Furthermore, the method further includes: constructing an observation probability matrix B using a probability density function; specifically, let the probability density function be , then the observation probability matrix B is expressed by the following formula:

[0092] ;

[0093] The probability density function is a multidimensional Gaussian probability density function, and is defined by the following formula:

[0094] ;

[0095] in, , represents the number of states of the hidden Markov chain, M represents the number of observations corresponding to each state, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as , Indicates the frame number.

[0096] It should be understood that the above Equivalent to the input of the probability density function, that is .

[0097] In this embodiment, , the number of states of the hidden Markov chain is 4, , the number of possible observations corresponding to each state is 3.

[0098] So far, the hidden Markov model parameter estimation method suitable for radar tracking data sequence classification has been clearly defined. A large amount of radar detection target tracking trajectory data is collected and the category information is manually labeled. Through the above description, the hidden Markov model parameters can be estimated. .

[0099] According to one aspect of the present invention, a method for object classification is provided.

[0100] Figure 2 This is a flow chart of a target classification method provided by an embodiment of the present invention.

[0101] like Figure 2 As shown, an object classification method provided by an embodiment of the present invention includes the following steps:

[0102] Step S21, obtaining tracking trajectory data of a target detected by the radar and generating a tracking data sequence, and extracting at least one feature vector corresponding to the target and used to characterize the target from the tracking data sequence to construct an observation data sequence;

[0103] Step S22: For each target, inputting the observation data sequence corresponding to the target into a hidden Markov model, and using the hidden Markov model to decode the tracking trajectory data of the target detected by the radar to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states;

[0104] Step S23, performing a logarithmic operation on the first target probability function of each state to obtain a second target probability function, and constructing a classifier function based on a portion of the second target probability function;

[0105] Step S24: determining the category of the current tracking trajectory data of the radar according to the state value corresponding to the minimum value output by the classifier function.

[0106] Specifically, according to the hidden Markov model modeled in the above embodiment, after obtaining the radar tracking data sequence, the RCS value, xSize value, and ySize value of the radar tracking data sequence are extracted to generate the observation data sequence , combined with the above hidden Markov model , you can start decoding the tracking trajectory data of the radar detection target to construct the first target probability function of the value of the observation data sequence corresponding to all different initial states. Because the hidden Markov model modeled in the above embodiment sets the transition probability matrix A to the unit matrix when modeling, that is, all hidden state values are equal. , the hidden state value at any time is the category of the tracking trajectory data of the radar.

[0107] For example, the original observation probability matrix is changed into the observation probability density matrix, and the observation value at time t is , using the probability density value of the current value to approximate its probability value.

[0108] Furthermore, for each target, the observation data sequence corresponding to the target is input into a hidden Markov model, and the tracking trajectory data of the target detected by the radar is decoded using the hidden Markov model to construct a first target probability function of the values of the observation data sequence corresponding to all different initial states, including:

[0109] Let the initial state , , represents the number of states of the hidden Markov chain, and the observation data sequence is expressed as hour,

[0110] Among them, among them, is a set of feature vectors, Expressed as [RCS ,xSize ,ySize ], , Indicates the frame number;

[0111] The transition probability matrix is set to an identity matrix to construct a first target probability function of the values of the observation data sequence corresponding to all different initial states.

[0112] Furthermore, the method of decoding the tracking trajectory data of the target detected by the radar using a hidden Markov model to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states includes:

[0113] The first target probability function of the value of the observed data sequence corresponding to all different initial states is calculated according to the following formula:

[0114] ;

[0115] ;

[0116] ;

[0117] in, represents the first target probability function of each state, represents the conditional probability, represents the observation data sequence, It indicates the probability of the observed data sequence X appearing when the model is in the hidden state H. is the initial state probability of the model, represents the probability density function, M represents the number of observations corresponding to each state, , represents the number of states of the hidden Markov chain, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0118] In an embodiment of the present invention, a unit matrix is used as a transfer probability matrix. Compared with the original decoding method which has a large amount of calculation and some calculation methods are also very complicated, it can simplify the model decoding operation and make the model decoding operation relatively simple.

[0119] However, in actual application, the above method is used to finally calculate the probability of the first target probability function under all different initial states. When the probability P is calculated, the computational overhead is high, and the calculated value is too small, which is not conducive to engineering applications. Therefore, the embodiments of the present invention mainly reduce the computational overhead of the probability P calculation and apply it to the real-time classification of millimeter-wave radar trajectory data to further reduce the computational workload. The following is an exemplary explanation.

[0120] According to the characteristics of the aforementioned first target probability function, the original formula of the first target probability function is mainly based on cumulative multiplication and exponential calculation, which itself has a great risk of numerical calculation. It is possible that the numerical calculation is very large or very small. For example, when the exponent is 100th power, 1000th power, 10000th power, or the exponent is -100th power or -1000th power, it is difficult to calculate using the computing power of existing computers. For this reason, the logarithm operation can be performed on the above-mentioned first target probability function to convert the cumulative multiplication and exponential calculation into cumulative operation, and the computational overhead can be greatly reduced.

[0121] Exemplarily, for a first target probability function of each state, performing a logarithmic operation on the first target probability function to obtain a second target probability function includes:

[0122] The first step is to , , represents the number of states of the hidden Markov chain, the second target probability function is obtained by the following formula:

[0123] ;

[0124] The second step is to process the second target probability function to obtain:

[0125] ;

[0126] The third step is to further process the second target probability function to obtain:

[0127] ;

[0128] Step 4: ,in,

[0129] ,in, is a constant term, ;

[0130] in, represents the second target probability function of each state, , represents the number of states of the hidden Markov chain, is the initial state probability of the model, represents the number of frames, M represents the number of observations corresponding to each state, represents the observation data sequence, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0131] So far, the values of the second target probability function corresponding to all different initial states are calculated , find the state value with the maximum probability , which is the category of the tracking trajectory data of the current radar detection target.

[0132] Furthermore, constructing a classifier function based on a portion of the second target probability function includes:

[0133] make , where the classifier function is obtained by the following formula:

[0134] ;

[0135] in, Represents the classifier function for each state, is the initial state probability of the model, , represents the number of states of the hidden Markov chain, Indicates the number of frames, is the mean vector, M represents the number of observations corresponding to each state, , is the covariance matrix, yes The inverse matrix of the covariance matrix is defined as .

[0136] It can be seen that the method of the embodiment of the present invention performs an equivalent replacement for the decoding calculation of the hidden Markov model. That is, the first target probability function is equivalently replaced by the second target probability function, and a classifier function is constructed based on the part of the second target probability function. The task of calculating the maximum value of the first target probability function can be equivalently replaced by calculating the minimum value of the classifier function. After the replacement, the original cumulative logarithmic calculation is replaced by a constant term plus cumulative multiplication, in which the cumulative multiplication is replaced by accumulation, and a large amount of exponential calculation is omitted, and the amount of calculation is greatly reduced. The original cumulative multiplication and exponential calculation result in very unstable values in the calculation process, and it is easy to have values very close to 0, which is difficult to calculate. After the equivalent replacement, only conventional calculations are required and the data are also relatively conventional values, which are suitable for direct calculation by a computer.

[0137] Furthermore, the method further comprises:

[0138] subtracting a current frame of the classifier function from a previous frame of the classifier function to construct a value-added function between two adjacent frames of the classifier function;

[0139] Calculating the values of the classifier function in the previous frame for all different initial states and caching them as reference values for the current frame;

[0140] The reference value and the value calculated by the value-added function are added to output the value of the classifier function in real time.

[0141] Furthermore, the method further comprises:

[0142] The category of the current tracking trajectory data of the radar is determined according to the state value corresponding to the minimum value of the current frame output by the classifier function.

[0143] In an embodiment of the present invention, compared with collecting all the observation values in the hidden Markov chain and then classifying the observation data sequence, it is possible to calculate and update the target category information in real time during the radar tracking target trajectory without having to wait until the target tracking trajectory is completed before classification.

[0144] Furthermore, the method further comprises:

[0145] The values of the current frame of the classifier function in all different initial states are cached and updated as the reference value of the next frame of the classifier function.

[0146] Specifically, the previous frame of the classifier function is defined as , the previous frame of the classifier function is represented by the following formula:

[0147] ;

[0148] The current frame of the classifier function is defined as , the current frame of the classifier function is represented by:

[0149] ;

[0150] right Processing yields:

[0151] ;

[0152] This is converted to:

[0153] ;

[0154] The value-added function is defined as ,but ,

[0155] ;

[0156] in, represents the value-added function, , represents the number of states of the hidden Markov chain, is the initial state probability of the model, Indicates the number of frames, represents the observation data sequence, represents the observation data sequence of the first frame, is the mean vector, , is the covariance matrix, yes The inverse matrix of yes The determinant of the covariance matrix is defined as .

[0157] As demonstrated by the above formula, as the radar trajectory data observation data sequence changes from time T Increase to T+1 When the classification equivalent alternative value An equivalent substitute value from the previous moment can be used Fast calculation, only needs to calculate the value of the value-added function and the baseline value of the previous frame.

[0158] Considering the characteristics of millimeter-wave radar trajectory data generation, when the target is first detected by the radar, the radar will keep tracking the current frame and generate observation data at the current time T every certain time (for example, 100ms). When classifying millimeter-wave radar trajectory data in real time, it is necessary to calculate the probability of the hidden Markov model classification for the radar trajectory data at each time T and find the maximum value. The technical solution of the embodiment of the present invention shows that when performing classification at time T+1, it is only necessary to calculate the incremental value of the value-added function. This is done by adding the incremental value to the equivalent replacement value (baseline value) of the previous frame T and finding the minimum value among the equivalent replacement values of the current frame T+1.

[0159] Thus, the target classification method provided by the embodiments of the present invention enables real-time classification of radar trajectory data using a hidden Markov model. Using an equivalent replacement scheme, this method simply calculates the increment of the current frame compared to the previous frame, adds this increment to the equivalent replacement value (baseline value) of the previous frame, and then finds the minimum value among the equivalent replacement values of the current frame. Compared to target classification methods that calculate equivalent replacement values at each time step, this method further reduces the computational effort, simplifying the original summation calculation to a single calculation.

[0160] According to another aspect of the present invention, a device for classifying an object is provided.

[0161] Figure 3 This is a structural block diagram of a target classification device provided by an embodiment of the present invention.

[0162] like Figure 3 As shown, the device 300 includes:

[0163] An acquisition module 310 is configured to acquire tracking trajectory data of a target detected by the radar and generate a tracking data sequence, and to extract at least one set of feature vectors corresponding to the target from the tracking data sequence to construct an observation data sequence;

[0164] A decoding module 320 is configured to input, for each target, an observation data sequence corresponding to the target into a hidden Markov model, and decode the tracking trajectory data of the target detected by the radar using the hidden Markov model to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states;

[0165] a classifier function construction module 330 for performing a logarithmic operation on the first target probability function for each state to obtain a second target probability function, and constructing a classifier function based on a portion of the second target probability function;

[0166] The determination module 340 is configured to determine the category of the current tracking trajectory data of the radar according to the state value corresponding to the minimum value output by the classifier function.

[0167] It should be understood that the execution principles, other aspects and effects of each unit (module) in the target classification device can be found in the contents of the aforementioned embodiments and will not be repeated here.

[0168] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the target classification method of any of the above-described embodiments is implemented.

[0169] The specific definition and implementation of the above steps can be found in the steps and methods of the embodiment of the radar target classification method, which will not be repeated here.

[0170] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0171] The beneficial effects of the present invention are as follows: compared with the existing technology, the target classification method, device and storage medium provided by the present invention solve the problems that radar in existing technology has difficulty in classifying detected targets and the low accuracy of target classification. By utilizing the tracking data sequence of each target within the radar monitoring range for classification, all information captured by the radar during the historical process of the target can be fully utilized, thereby greatly improving the accuracy of radar target classification.

[0172] Furthermore, the Hidden Markov Model (HMM) was simplified, reducing the computational complexity of model decoding. A continuous HMM approach was used to adapt to the characteristics of millimeter-wave radar data and to classify radar target trajectory data. Using the HMM, the classification of traffic millimeter-wave radar data was very effective, with a classification recall rate exceeding 95%.

[0173] Furthermore, the embodiments of the present invention significantly reduce the computational overhead of classifying radar trajectory data using a hidden Markov model.

[0174] Furthermore, the embodiment of the present invention can calculate and update the target category information in real time during the process of radar tracking the target trajectory, without having to wait until the target tracking trajectory is completed before classification.

[0175] The above is a detailed introduction to a target classification method, device and storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A target classification method, characterized in that: The method comprises: Acquiring tracking trajectory data of a target detected by the radar and generating a tracking data sequence, and extracting at least one feature vector corresponding to the target and used to characterize the target from the tracking data sequence to construct an observation data sequence; For each target, inputting the observation data sequence corresponding to the target into a hidden Markov model, and using the hidden Markov model to decode the tracking trajectory data of the target detected by the radar to obtain a first target probability function of the value of the observation data sequence corresponding to all different initial states; For each target, the observation data sequence corresponding to the target is input into the hidden Markov model, and the tracking trajectory data of the target detected by the radar is decoded using the hidden Markov model to obtain the first target probability function of the value of the observation data sequence corresponding to all different initial states. Let the initial state h1 = S i ,i∈[1,...,N], N represents the number of states of the hidden Markov chain, S i Represents the hidden state, and the observation data sequence is X=[X1,X2,…,X q ,…,X T ]hour, Among them, X q is a set of eigenvectors, X q Expressed as [RCSq, xSizeq, ySizeq], q∈[1,...,T], T represents the number of frames; Setting the transition probability matrix in the model to the identity matrix to construct a first target probability function of the values of the observation data sequence corresponding to all different initial states; For the first target probability function of each state, performing a logarithmic operation on the first target probability function to obtain a second target probability function, and constructing a classifier function based on a portion of the second target probability function; constructing a classifier function based on a portion of the second target probability function includes: setting Among them, the classifier function is obtained by the following formula: in, Represents the classifier function of each state, π i is the initial state probability of the model, i∈[1,…,N], T represents the number of frames, μ=[μ1,μ2,…,μ M ] is the mean vector, M represents the number of observations corresponding to each state, μ=E[O], Σ i is the covariance matrix, Σ i -1 is Σ i The inverse matrix of |Σ i |YesΣ i The determinant, covariance matrix is defined as ∑ i =E[(x-μ)(x-μ) T ]; The category of the current tracking trajectory data of the radar is determined according to the state value corresponding to the minimum value output by the classifier function.

2. The target classification method according to claim 1, wherein: The method of decoding the tracking trajectory data of the detection target of the radar by using a hidden Markov model to obtain a first target probability function of the value of the observation data sequence corresponding to all different initial states includes: The first target probability function of the value of the observed data sequence corresponding to all different initial states is calculated according to the following formula: Among them, p i represents the first target probability function of each state, represents the conditional probability, X represents the observed data sequence, Indicates the probability of the observed data sequence X when the model is in the hidden state H, G i (O=x j ) represents the probability density function.

3. The target classification method according to claim 2, wherein: For each state, the method of performing a logarithmic operation on the first target probability function to obtain a second target probability function includes: The first step is to i∈[1,…,N], N represents the number of states of the hidden Markov chain, and the second target probability function is obtained by the following formula: The second step is to process the second target probability function to obtain: The third step is to further process the second target probability function to obtain: Step 4: in, Among them, c is a constant term, Where P is the probability value of the first target probability function, is the value of the second target probability function, represents the second target probability function of each state, and X represents the observation data sequence.

4. The target classification method according to claim 1, wherein: The method further comprises: subtracting a current frame of the classifier function from a previous frame of the classifier function to construct a value-added function between two adjacent frames of the classifier function; Calculating the values of the classifier function in the previous frame for all different initial states and caching them as reference values for the current frame; The reference value and the value calculated by the value-added function are added to output the value of the classifier function in real time.

5. The target classification method according to claim 4, wherein: The method further comprises: The category of the current tracking trajectory data of the radar is determined according to the state value corresponding to the minimum value of the current frame output by the classifier function.

6. The target classification method according to claim 5, wherein: The method further comprises: The values of the current frame of the classifier function in all different initial states are cached and updated as the reference value of the next frame of the classifier function.

7. The target classification method according to claim 4, wherein: The method of subtracting the current frame of the classifier function from the previous frame of the classifier function to construct a value-added function between two adjacent frames of the classifier function includes: The previous frame of the classifier function is defined as The previous frame of the classifier function is represented by: The current frame of the classifier function is defined as The current frame of the classifier function is represented by: right Processing yields: This is converted to: The value-added function is defined as but in, Represents the value-added function, X represents the observation data sequence, and x1 represents the observation data sequence of the first frame.

8. A target classification device, characterized in that: The device comprises: an acquisition module, configured to acquire tracking trajectory data of a target detected by the radar and generate a tracking data sequence, and extract at least one feature vector corresponding to the target from the tracking data sequence to construct an observation data sequence; A decoding module is configured to input, for each target, an observation data sequence corresponding to the target into a hidden Markov model, and use the hidden Markov model to decode the tracking trajectory data of the target detected by the radar to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states; the inputting, for each target, an observation data sequence corresponding to the target into the hidden Markov model, and use the hidden Markov model to decode the tracking trajectory data of the target detected by the radar to obtain a first target probability function of the values of the observation data sequence corresponding to all different initial states includes: Let the initial state h1 = S i ,i∈[1,...,N], N represents the number of states of the hidden Markov chain, S i Represents the hidden state, and the observation data sequence is X=[X1,X2,…,X q ,…,X T ]hour, Among them, X q is a set of eigenvectors, X q Expressed as [RCSq, xSizeq, ySizeq], q∈[1,...,T], T represents the number of frames; Setting the transition probability matrix in the model to the identity matrix to construct a first target probability function of the values of the observation data sequence corresponding to all different initial states; A classifier function construction module is used to perform a logarithmic operation on the first target probability function of each state to obtain a second target probability function, and construct a classifier function based on a portion of the second target probability function; the construction of the classifier function based on the portion of the second target probability function includes: setting Among them, the classifier function is obtained by the following formula: in, Represents the classifier function of each state, π i is the initial state probability of the model, i∈[1,…,N], T represents the number of frames, μ=[μ1,μ2,…,μ M ] is the mean vector, M represents the number of observations corresponding to each state, μ=E[O], Σ i is the covariance matrix, Σ i -1 is Σ i The inverse matrix of |Σ i |YesΣ i The determinant, covariance matrix is defined as ∑ i =E[(x-μ)(x-μ) T ]; The determination module is used to determine the category of the current tracking trajectory data of the radar according to the state value corresponding to the minimum value output by the classifier function.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the target classification method according to any one of claims 1 to 7 is implemented.

Images