CN112014821A - Unknown vehicle target identification method based on radar broadband characteristics - Google Patents
Unknown vehicle target identification method based on radar broadband characteristics Download PDFInfo
- Publication number
- CN112014821A CN112014821A CN202010880964.5A CN202010880964A CN112014821A CN 112014821 A CN112014821 A CN 112014821A CN 202010880964 A CN202010880964 A CN 202010880964A CN 112014821 A CN112014821 A CN 112014821A
- Authority
- CN
- China
- Prior art keywords
- sample
- dimension
- unknown
- sample set
- detected
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/411—Identification of targets based on measurements of radar reflectivity
- G01S7/412—Identification of targets based on measurements of radar reflectivity based on a comparison between measured values and known or stored values
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Abstract
The invention provides an unknown vehicle target identification method based on radar broadband characteristics, and belongs to the technical field of unknown vehicle target identification. Firstly, taking radar one-dimensional range profiles of a known vehicle type and an unknown vehicle type as a sample set to be measured; encoding a sample set to be detected by using a self-encoder; calculating to obtain a division threshold of the known vehicle type and the unknown vehicle type by using an anomaly detection method; calculating to obtain the discrimination probability of each sample in the sample set to be detected; and comparing the discrimination probability with a threshold, and rejecting a sample set to be detected according to a comparison result. The method is based on the one-dimensional radar high-resolution range profile data which is simple in data structure and can be provided all weather, and the self-encoder, the error sorting and the anomaly detection algorithm are combined to be used for identifying and eliminating the unknown vehicle types, so that the method is prepared for the subsequent training of the known type identification network.
Description
Technical Field
The invention belongs to the technical field of unknown vehicle target identification, and particularly relates to an unknown vehicle target identification method based on radar broadband characteristics.
Background
The current research of the vehicle unknown identification method is based on the research of taking videos and pictures as data sampling equipment and taking a machine learning method as a technical basis. The method is characterized in that vehicle data of unknown classes except a vehicle set to be classified are resampled and collected, input into the whole generation countermeasure network (comprising a generator and a discriminator) to be trained, and meanwhile, vehicle video or picture data of known classes are jointly used for training the discriminator network of the generation countermeasure network, so that the network classification capability is improved, the network screening weight capable of discriminating the vehicles of the unknown classes at a high probability is saved, and finally generated discriminator network weights can identify the vehicles of different classes and find out vehicle samples of the unknown classes. However, the drawbacks of the background art are mainly reflected in two aspects:
(1) the data collected by the background technology has the defects of large batch data quantity and long time consumption in the data preprocessing process. The video or picture data itself carries a large amount of data, and a two-dimensional high-definition picture (a sample of data) has a size of about 10 after preprocessing2KB or so, and the number of data set samples for deep learning network training must be kept at 105About the order of magnitude. It is expected that the method for identifying the feature data by using the video or the picture consumes long time and much resources from data preprocessing to identification. And the picture and video data can only be provided during the daytime when the weather conditions are good, but cannot be provided at night or under the condition of complex weather.
(2) The generation of a countermeasure network (GAN) used in the background art requires a large number of data samples to identify itself, and in order to achieve the purpose of identifying an unknown vehicle object, sample data of a plurality of unknown vehicles (i.e., no tags) needs to be collected randomly on a road, and then the network is trained by combining the sample data of the known vehicles. The process of collecting sample data of unknown types and training can further lengthen the running time of the algorithm and increase the consumption of operation resources.
Disclosure of Invention
Aiming at the defects in the prior art, the unknown vehicle target identification method based on the radar broadband characteristics provided by the invention solves the problems.
In order to achieve the above purpose, the invention adopts the technical scheme that:
the scheme provides an unknown vehicle target identification method based on radar broadband characteristics, which comprises the following steps:
s1, respectively acquiring radar one-dimensional range profiles of a known vehicle type and an unknown vehicle type, taking the radar one-dimensional range profiles of the known vehicle type and the unknown vehicle type as a sample set to be detected, and respectively labeling a known sample to be detected and an unknown sample to be detected;
s2, coding the sample set to be detected by using the self-coder to obtain a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected;
s3, respectively setting ideal harmonic mean values, initial threshold values, threshold probability step values and total iteration times corresponding to the precision ratio and the recall ratio of the sample set to be tested;
s4, obtaining a first division threshold of a known vehicle type sample and an unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value, and calculating to obtain the discrimination probability of each sample in the sample set to be detected;
s5, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold according to the channel, and calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the judgment result;
s6, updating the first division threshold according to the actual harmonic mean value to obtain a current threshold;
s7, judging whether the actual harmonic mean value is larger than or equal to a preset ideal harmonic mean value, if so, outputting a current threshold, and entering a step S9, otherwise, entering a step S8;
s8, judging whether the iteration frequency reaches the total iteration frequency, if so, selecting a threshold corresponding to an ideal harmonic mean value or a threshold corresponding to a historical maximum harmonic mean value, obtaining a second division threshold of a known vehicle type sample and an unknown vehicle type sample according to the threshold corresponding to the ideal harmonic mean value or the threshold corresponding to the historical maximum harmonic mean value, and entering the step S9, otherwise, returning to the step S4;
and S9, judging whether the current threshold or the second division threshold is larger than the judgment probability of the actual measurement sample, if so, marking the current threshold or the second division threshold as an unknown vehicle type, otherwise, marking the current threshold or the second division threshold as a known vehicle type, and completing the identification of the unknown vehicle target.
The invention has the beneficial effects that: the invention is based on the one-dimensional radar high-resolution range profile data which has a simpler data structure and can be provided all weather to carry out operation, and combines and uses an auto-encoder (AE), an error sorting and an anomaly detection algorithm to identify and eliminate the unknown vehicle category, thereby preparing for the subsequent training of a known category identification network. Meanwhile, labels are marked on the known vehicle samples and the unknown vehicle samples in the sample set, so that iterative computation of the F1Score harmonic mean value is facilitated.
Further, the step S2 includes the following steps:
s201, taking the radar one-dimensional range profile of the known vehicle category as a training set, and training a self-encoder by using the training set;
s202, reconstructing the sample set to be tested by utilizing the trained self-encoder to obtain the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s203, screening out the dimension with the minimum error in the known vehicle category sample and the dimension with the maximum error in the unknown vehicle category sample according to the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s204, sequencing the average reconstruction errors of each dimension in the known vehicle category samples in a descending order, and sequencing the average reconstruction errors of each dimension in the unknown vehicle category samples in a descending order;
s205, according to the sequencing result, extracting the intersection of the dimension set of the known vehicle class sample and the unknown vehicle class sample in the first m dimensions, and obtaining a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected according to the intersection.
The beneficial effects of the further scheme are as follows: the method uses the radar one-dimensional range profile as sample data, improves the usability of the data and the robustness of an application scene, and uses the self-encoder to extract errors, so that the speed is obviously higher than that of joint training of a generator and a discriminator for generating a countermeasure network; the screening and sorting of the channels with the minimum error and the maximum error can greatly reduce the identification failure rate of unknown classes, so that the identification precision and the time consumption can be controlled by the number of sorting channels.
Still further, the expression of the average reconstruction error of each dimension in the known vehicle category samples in step S202 is as follows:
wherein, Average (1 j) Represents the average reconstruction error of the jth dimension in the known vehicle class samples, and j belongs to {1,2, 3.., M }, M represents the dimension total number of each sample in the sample set to be tested,1 jrepresenting the total reconstruction error, N, in a sample of known vehicle classes1A total number of one-dimensional range profile training data representing a known vehicle class;
the expression for the average reconstruction error for each dimension in the unknown vehicle category sample is as follows:
wherein the content of the first and second substances,represents the average reconstruction error in the j dimension of the unknown vehicle class sample,representing total reconstruction errors in unknown vehicle class samples,N2A total number of one-dimensional range profile training data representing unknown vehicle classes.
The beneficial effects of the further scheme are as follows: the calculation result of the self-encoder has certain randomness, and the randomness can be well controlled by averaging the reconstruction error of each channel, so that the error difference of the known class and the unknown class in different channels in the sample set to be detected can be more objectively reflected.
Still further, in step S205, an expression of the intersection of the dimension sets of the known vehicle category sample and the unknown vehicle category sample is as follows:
j∩=j1∩j2
j1∈{1,2,...,m1}
j2∈{1,2,...,m2}
wherein j is∩Representing the intersection of the set of dimensions, j, of the known vehicle class sample and the unknown vehicle class sample1Represents the front m1Dimension, jj, that minimizes the average reconstruction error for the jth dimension in the known vehicle class2Represents the front m2And the dimension which maximizes the average reconstruction error of the jth dimension in the unknown vehicle class.
The beneficial effects of the further scheme are as follows: the channels with the smallest known class error and the largest unknown class error can be found out to the maximum extent by solving the intersection of the dimension sets, and on the channels, the overall distribution of the samples can be distinct, namely, the errors of most of the known class samples are smaller, and the errors of the other unknown class samples are larger, so that the difference of Gaussian distribution obtained based on the error vectors is obvious, and the judgment threshold can be obtained more conveniently.
The beneficial effects of the further scheme are as follows: the invention combines the self-encoder with the abnormity detection, thereby saving a huge training process for generating the countermeasure network and improving the identification performance for identifying the unknown vehicle category.
Still further, the step S4 includes the steps of:
s401, obtaining a first division threshold of a known vehicle type sample and an unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value;
s402, taking the intersection dimensionality of the known vehicle type sample and the unknown vehicle type sample as the dimensionality of a sample set to be detected;
s403, calculating to obtain a mean value of each dimension in the sample set to be detected aiming at each dimension in the sample set to be detected, and calculating to obtain a variance estimation value of each dimension in the sample set to be detected according to the mean value;
s404, calculating to obtain the discrimination probability of each sample in the sample set to be detected according to the mean value and variance estimation value of each dimension in the sample set to be detected.
The beneficial effects of the further scheme are as follows: the method lays a foundation for judging the unknown vehicle target type by calculating the discrimination rate.
Still further, in step S402, an expression of the mean value of each dimension in the sample set to be measured is as follows:
wherein, mujRepresents the mean value of each dimension in the sample set to be tested, m represents the total number of samples to be tested,the method comprises the steps that a value of the jth dimension of a sample is represented, i represents a sample number, i belongs to {1,2, 3., M }, j represents the number of the dimensions, j belongs to {1,2, 3., M }, and M represents the total number of the dimensions of each sample in a sample set to be tested;
the expression of the variance estimation value of each dimension in the sample set to be tested is as follows:
wherein the content of the first and second substances,and representing the variance estimation value of each dimension in the sample set to be measured.
The beneficial effects of the further scheme are as follows: the mean and variance estimates for each dimension are calculated to provide for the next step of gaussian distribution probability.
Still further, the expression of the discrimination probability of each sample in the sample set to be measured in step S403 is as follows:
wherein the content of the first and second substances,the method includes the steps that the discrimination probability of each sample in a sample set to be detected is represented, j represents the number of dimensions, j belongs to {1,2, 3.., M }, M represents the total number of dimensions of each sample in the sample set to be detected, and sigma represents the total number of dimensions of each sample in the sample set to be detectedjRepresents the standard deviation of the jth dimension,represents the value of the j-th dimension in the sample, μjThe mean value of each of the dimensions is represented,and representing the variance estimation value of each dimension of the sample set to be measured.
The beneficial effects of the further scheme are as follows: the invention utilizes the comparison between the discrimination rate and the threshold to identify and eliminate the unknown vehicle types.
Still further, the step S5 includes the steps of:
s501, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold or not according to the channel;
s502, calculating precision ratio and recall ratio of the known vehicle type sample and the unknown vehicle type sample by using an anomaly detection method;
and S502, calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the precision ratio and the recall ratio.
The beneficial effects of the further scheme are as follows: the invention combines the self-encoder with the abnormity detection, thereby saving a huge training process for generating the countermeasure network and improving the identification performance for identifying the unknown vehicle category.
Still further, the expression of the actual harmonic mean in step S502 is as follows:
wherein, F1Score represents the actual harmonic mean, P represents the precision of the known and unknown vehicle class samples, R represents the recall of the known and unknown vehicle class samples, TP represents the positive sample of a correct prediction, FN represents the negative sample of a wrong prediction, and FP represents the positive sample of a wrong prediction.
The beneficial effects of the further scheme are as follows: the judgment threshold of the unknown target is determined by calculating the actual harmonic mean value, and the distribution of the test signals under the real condition is assumed to be consistent with the distribution of the sample set to be tested, so that the threshold obtained by iterative calculation can be used as a basis for removing real unknown samples to analyze and remove unknown samples in actual measurement samples under the future real application scene.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
Examples
In the actual identification application, the unknown type of vehicle target is mixed into the detected target set to cause identification errors, and the target set of the unknown type of target is removed before the vehicle target is identified. As shown in fig. 1, the present invention provides an unknown vehicle target identification method based on radar broadband characteristics, which is implemented as follows:
s1, respectively acquiring radar one-dimensional range profiles of a known vehicle type and an unknown vehicle type, taking the radar one-dimensional range profiles of the known vehicle type and the unknown vehicle type as a sample set to be detected, and respectively labeling a known sample to be detected and an unknown sample to be detected;
in this embodiment, the radar one-dimensional range profile of the known vehicle category and the radar one-dimensional range profile of the unknown vehicle category are used as a sample set to be measured and divided into a training set and a verification set, the training set is used for subsequent training of the self-encoder, and the verification set is used for obtaining a threshold.
In this embodiment, the known sample to be tested and the unknown sample to be tested are labeled respectively, so that the harmonic mean F1Score can be calculated conveniently in the subsequent iteration. The algorithm is that the distribution of the test signals of the real situation is assumed to be consistent with the distribution of the sample set to be tested, so that the threshold e obtained by iterative calculation of the sample set to be tested can be used as the basis for removing real unknown samples, and samples of unknown types in the actual measurement samples in the future real application scene are removed by analysis and removal.
S2, coding the sample set to be detected by using the self-coder to obtain a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected, wherein the implementation method comprises the following steps:
s201, taking the radar one-dimensional range profile of the known vehicle category as a training set, and training a self-encoder by using the training set;
s202, reconstructing the sample set to be tested by utilizing the trained self-encoder to obtain the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s203, screening out the dimension with the minimum error in the known vehicle category sample and the dimension with the maximum error in the unknown vehicle category sample according to the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s204, sequencing the average reconstruction errors of each dimension in the known vehicle category samples in a descending order, and sequencing the average reconstruction errors of each dimension in the unknown vehicle category samples in a descending order;
s205, according to the sequencing result, extracting the intersection of the dimension set of the known vehicle class sample and the unknown vehicle class sample in the first m dimensions, and obtaining a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected according to the intersection.
In this embodiment, assuming that the dimensions of the high-resolution one-dimensional range profile (HRRP) of the vehicle target are independent, a stacked self-encoder (SAE) is applied to encode and decode the HRRP corresponding to each target in the original detection target set. Assuming known class HRRP training set data asHRRP training set data of unknown classUse ofTraining the self-encoder and using it respectivelyAndobtaining a reconstructed HRRP through a trained self-encoder, and calculating the average reconstruction error of each dimension of different training setsj(assuming that the HRRP dimension is j epsilon {1,2, 3.,. M }), and screening out dimensions which simultaneously meet the conditions that the average reconstruction error of the known class is small and the average reconstruction error of the unknown class is large. The average reconstruction error for the jth dimension is defined as:
wherein the content of the first and second substances,representing the average reconstruction error in the j dimension of the sample of known vehicle classes,representing the total reconstruction error, N, in a sample of known vehicle classes1The total number of one-dimensional range profile training data representing known vehicle classes,represents the average reconstruction error in the j dimension of the unknown vehicle class sample,representing the total reconstruction error, N, in a sample of unknown vehicle classes2A total number of one-dimensional range profile training data representing unknown vehicle classes.
Respectively at y1And y2M dimensions of1Are all made ofSmallest dimension j1∈{1,2,...,m1And m, and m2Are all made ofMaximum dimension j2∈{1,2,...,m2}。
At j1And j2And (3) obtaining dimension intersection:
j∩=j1∩j2
wherein j is∩Representing the intersection of the set of dimensions, j, of the known vehicle class sample and the unknown vehicle class sample1Represents the front m1Dimension, jj, that minimizes the average reconstruction error for the jth dimension in the known vehicle class2Represents the front m2The dimension that minimizes the average reconstruction error for the jth dimension in the unknown vehicle class.
S3, respectively setting ideal harmonic mean values, initial threshold values, threshold probability step values and total iteration times corresponding to the precision ratio and the recall ratio of the sample set to be tested;
in this embodiment, the threshold satisfying the ideal harmonic mean can be found by effectively iterating the threshold values.
S4, obtaining a first division threshold of the known vehicle type sample and the unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value, and calculating to obtain the discrimination probability of each sample in the sample set to be detected, wherein the implementation method comprises the following steps:
s401, obtaining a first division threshold of a known vehicle type sample and an unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value;
s402, taking the intersection dimensionality of the known vehicle type sample and the unknown vehicle type sample as the dimensionality of a sample set to be detected;
s403, calculating to obtain a mean value of each dimension in the sample set to be detected aiming at each dimension in the sample set to be detected, and calculating to obtain a variance estimation value of each dimension in the sample set to be detected according to the mean value;
s404, calculating to obtain the discrimination probability of each sample in the sample set to be detected according to the mean value and variance estimation value of each dimension in the sample set to be detected.
In this embodiment, the verification set is constructed by using the sample to be tested(wherein, M samples in total, each sample is M-dimensional) which contains HRRP samples corresponding to the vehicle targets of the known class and the vehicle targets of the unknown class, and the dimensions of all the samples are taken as j∩Of (2). Assuming that the samples in the verification set do not satisfy the Gaussian distribution, the original data set needs to be subjected toTaking logarithm:
wherein x represents an adjustable parameter, yvRepresenting a log-taken validation set.
Finally, theCalculating the discrimination probability of the data to be measuredAnd make it satisfyIt is determined to be an unknown class vehicle target, where it represents a discrimination threshold.
S5, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold according to the channel, and calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the judgment result, wherein the realization method comprises the following steps:
s501, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold or not according to the channel;
s502, calculating precision ratio and recall ratio of the known vehicle type sample and the unknown vehicle type sample by using an anomaly detection method;
and S502, calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the precision ratio and the recall ratio.
In this embodiment, in order to determine the decision threshold of the unknown class object, precision rate (precision) and recall rate (call) are used as an auxiliary. It is known that TP represents the positive samples of correct prediction, FN represents the negative samples of incorrect prediction, and FP represents the positive samples of incorrect prediction. The precision ratio P and the recall ratio R are respectively defined as:
let the harmonic mean of P and R be F1Score, then its expression is:
in this embodiment, the determination threshold of the unknown target is determined by calculating the actual harmonic mean value, and it is assumed that the distribution of the test signal in the real situation is consistent with the distribution of the sample set to be tested, so that the threshold obtained by iterative computation can be used as a basis for removing the real unknown sample to analyze and remove the unknown sample in the actual measurement sample in the future real application scene.
S6, updating the first division threshold according to the actual harmonic mean value to obtain a current threshold;
s7, judging whether the actual harmonic mean value is larger than or equal to a preset ideal harmonic mean value, if so, outputting a current threshold, and entering a step S9, otherwise, entering a step S8;
s8, judging whether the iteration frequency reaches the total iteration frequency, if so, selecting a threshold corresponding to an ideal harmonic mean value or a threshold corresponding to a historical maximum harmonic mean value, obtaining a second division threshold of a known vehicle type sample and an unknown vehicle type sample according to the threshold corresponding to the ideal harmonic mean value or the threshold corresponding to the historical maximum harmonic mean value, and entering the step S9, otherwise, returning to the step S4;
and S9, judging whether the current threshold or the second division threshold is larger than the judgment probability of the actual measurement sample, if so, marking the current threshold or the second division threshold as an unknown vehicle type, otherwise, marking the current threshold or the second division threshold as a known vehicle type, and completing the identification of the unknown vehicle target.
In this embodiment, the sample to be tested is actually usedIs determined probability valueThe calculation is carried out and compared with the calculation,the samples are marked as unknown classes and the sample set to be tested is rejected.
The invention is based on the one-dimensional radar high-resolution range profile data which has a simpler data structure and can be provided all weather to carry out operation, and combines and uses an auto-encoder (AE), an error sorting and an anomaly detection algorithm to identify and eliminate the unknown vehicle types, thereby preparing for the subsequent training of a known type identification network.
Claims (9)
1. An unknown vehicle target identification method based on radar broadband characteristics is characterized by comprising the following steps:
s1, respectively acquiring radar one-dimensional range profiles of a known vehicle type and an unknown vehicle type, taking the radar one-dimensional range profiles of the known vehicle type and the unknown vehicle type as a sample set to be detected, and respectively labeling a known sample to be detected and an unknown sample to be detected;
s2, coding the sample set to be detected by using the self-coder to obtain a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected;
s3, respectively setting ideal harmonic mean values, initial threshold values, threshold probability step values and total iteration times corresponding to the precision ratio and the recall ratio of the sample set to be tested;
s4, obtaining a first division threshold of a known vehicle type sample and an unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value, and calculating to obtain the discrimination probability of each sample in the sample set to be detected;
s5, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold according to the channel, and calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the judgment result;
s6, updating the first division threshold according to the actual harmonic mean value to obtain a current threshold;
s7, judging whether the actual harmonic mean value is larger than or equal to a preset ideal harmonic mean value, if so, outputting a current threshold, and entering a step S9, otherwise, entering a step S8;
s8, judging whether the iteration frequency reaches the total iteration frequency, if so, selecting a threshold corresponding to an ideal harmonic mean value or a threshold corresponding to a historical maximum harmonic mean value, obtaining a second division threshold of a known vehicle type sample and an unknown vehicle type sample according to the threshold corresponding to the ideal harmonic mean value or the threshold corresponding to the historical maximum harmonic mean value, and entering the step S9, otherwise, returning to the step S4;
and S9, judging whether the current threshold or the second division threshold is larger than the judgment probability of the actual measurement sample, if so, marking the current threshold or the second division threshold as an unknown vehicle type, otherwise, marking the current threshold or the second division threshold as a known vehicle type, and completing the identification of the unknown vehicle target.
2. The method for identifying an unknown vehicle target based on radar broadband characteristics as claimed in claim 1, wherein said step S2 includes the steps of:
s201, taking the radar one-dimensional range profile of the known vehicle category as a training set, and training a self-encoder by using the training set;
s202, reconstructing the sample set to be tested by utilizing the trained self-encoder to obtain the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s203, screening out the dimension with the minimum error in the known vehicle category sample and the dimension with the maximum error in the unknown vehicle category sample according to the average reconstruction error of each dimension in the known vehicle category sample and the average reconstruction error of each dimension in the unknown vehicle category sample;
s204, sequencing the average reconstruction errors of each dimension in the known vehicle category samples in a descending order, and sequencing the average reconstruction errors of each dimension in the unknown vehicle category samples in a descending order;
s205, according to the sequencing result, extracting the intersection of the dimension set of the known vehicle class sample and the unknown vehicle class sample in the first m dimensions, and obtaining a channel which meets the minimum reconstruction error of the known vehicle class and the maximum reconstruction error of the unknown vehicle class in the sample set to be detected according to the intersection.
3. The unknown vehicle target identification method based on the radar broadband characteristics as claimed in claim 2, wherein the expression of the average reconstruction error of each dimension in the known vehicle class samples in the step S202 is as follows:
wherein the content of the first and second substances,represents the average reconstruction error of the jth dimension in the known vehicle class samples, and j belongs to {1,2, 3.., M }, M represents the dimension total number of each sample in the sample set to be tested,representing the total reconstruction error, N, in a sample of known vehicle classes1A total number of one-dimensional range profile training data representing a known vehicle class;
the expression for the average reconstruction error for each dimension in the unknown vehicle category sample is as follows:
wherein the content of the first and second substances,represents the average reconstruction error in the j dimension of the unknown vehicle class sample,representing the total reconstruction error, N, in a sample of unknown vehicle classes2A total number of one-dimensional range profile training data representing unknown vehicle classes.
4. The unknown vehicle target identification method based on the radar broadband characteristic of claim 2, wherein the expression of the intersection of the dimension sets of the known vehicle class samples and the unknown vehicle class samples in the step S205 is as follows:
j∩=j1∩j2
j1∈{1,2,...,m1}
j2∈{1,2,...,m2}
wherein j is∩Representing the intersection of the set of dimensions, j, of the known vehicle class sample and the unknown vehicle class sample1Represents the front m1Dimension, jj, that minimizes the average reconstruction error for the jth dimension in the known vehicle class2Represents the front m2And the dimension which maximizes the average reconstruction error of the jth dimension in the unknown vehicle class.
5. The unknown vehicle target identification method based on radar broadband characteristics as claimed in claim 2, wherein said step S4 includes the steps of:
s401, obtaining a first division threshold of a known vehicle type sample and an unknown vehicle type sample in a single cycle according to the threshold probability step value and the initial threshold value;
s402, taking the intersection dimensionality of the known vehicle type sample and the unknown vehicle type sample as the dimensionality of a sample set to be detected;
s403, calculating to obtain a mean value of each dimension in the sample set to be detected aiming at each dimension in the sample set to be detected, and calculating to obtain a variance estimation value of each dimension in the sample set to be detected according to the mean value;
s404, calculating to obtain the discrimination probability of each sample in the sample set to be detected according to the mean value and variance estimation value of each dimension in the sample set to be detected.
6. The method for identifying an unknown vehicle target based on radar broadband characteristics as claimed in claim 5, wherein the expression of the mean value of each dimension in the sample set to be tested in step S402 is as follows:
wherein, mujRepresents the mean value of each dimension in the sample set to be tested, m represents the total number of samples to be tested,the method comprises the steps that a value of the jth dimension of a sample is represented, i represents a sample number, i belongs to {1,2, 3., M }, j represents the number of the dimensions, j belongs to {1,2, 3., M }, and M represents the total number of the dimensions of each sample in a sample set to be tested;
the expression of the variance estimation value of each dimension in the sample set to be tested is as follows:
7. The method for identifying an unknown vehicle target based on radar broadband characteristics as claimed in claim 5, wherein the expression of the discrimination probability of each sample in the sample set to be tested in the step S403 is as follows:
wherein the content of the first and second substances,the method includes the steps that the discrimination probability of each sample in a sample set to be detected is represented, j represents the number of dimensions, j belongs to {1,2, 3.., M }, M represents the total number of dimensions of each sample in the sample set to be detected, and sigma represents the total number of dimensions of each sample in the sample set to be detectedjRepresents the standard deviation of the jth dimension,represents the value of the j-th dimension in the sample, μjThe mean value of each of the dimensions is represented,and representing the variance estimation value of each dimension of the sample set to be measured.
8. The method for identifying an unknown vehicle target based on radar broadband characteristics as claimed in claim 1, wherein said step S5 includes the steps of:
s501, judging whether the judgment probability of each sample in the sample set to be detected is smaller than a first division threshold or not according to the channel;
s502, calculating precision ratio and recall ratio of the known vehicle type sample and the unknown vehicle type sample by using an anomaly detection method;
and S502, calculating to obtain the actual harmonic mean value of the precision ratio and the recall ratio of the sample set to be detected according to the precision ratio and the recall ratio.
9. The method for identifying an unknown vehicle target based on radar broadband characteristics as claimed in claim 8, wherein the expression of the actual harmonic mean in step S502 is as follows:
wherein, F1Score represents the actual harmonic mean, P represents the precision of the known and unknown vehicle class samples, R represents the recall of the known and unknown vehicle class samples, TP represents the positive sample of a correct prediction, FN represents the negative sample of a wrong prediction, and FP represents the positive sample of a wrong prediction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010880964.5A CN112014821B (en) | 2020-08-27 | 2020-08-27 | Unknown vehicle target identification method based on radar broadband characteristics |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010880964.5A CN112014821B (en) | 2020-08-27 | 2020-08-27 | Unknown vehicle target identification method based on radar broadband characteristics |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112014821A true CN112014821A (en) | 2020-12-01 |
CN112014821B CN112014821B (en) | 2022-05-17 |
Family
ID=73502528
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010880964.5A Active CN112014821B (en) | 2020-08-27 | 2020-08-27 | Unknown vehicle target identification method based on radar broadband characteristics |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112014821B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113240011A (en) * | 2021-05-14 | 2021-08-10 | 烟台海颐软件股份有限公司 | Deep learning driven abnormity identification and repair method and intelligent system |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110007355A (en) * | 2019-04-15 | 2019-07-12 | 中国科学院电子学研究所 | The detection method and device of a kind of convolution self-encoding encoder and interior of articles exception |
US20190317191A1 (en) * | 2018-04-11 | 2019-10-17 | Infineon Technologies Ag | Human Detection and Identification in a Setting Using Millimiter-Wave Radar |
US20190383904A1 (en) * | 2018-06-13 | 2019-12-19 | Metawave Corporation | Autoencoder assisted radar for target identification |
CN110658507A (en) * | 2019-10-12 | 2020-01-07 | 电子科技大学 | Multi-class average maximization true and false target feature extraction method for radar target identification |
CN110691100A (en) * | 2019-10-28 | 2020-01-14 | 中国科学技术大学 | Hierarchical network attack identification and unknown attack detection method based on deep learning |
CN111144462A (en) * | 2019-12-16 | 2020-05-12 | 中国电子科技集团公司第三十八研究所 | Unknown individual identification method and device for radar signals |
CN111222575A (en) * | 2020-01-07 | 2020-06-02 | 北京联合大学 | KLXS multi-model fusion method and system based on HRRP target recognition |
CN111273288A (en) * | 2020-03-06 | 2020-06-12 | 电子科技大学 | Radar unknown target identification method based on long-term and short-term memory network |
CN111352086A (en) * | 2020-03-06 | 2020-06-30 | 电子科技大学 | Unknown target identification method based on deep convolutional neural network |
US20210033724A1 (en) * | 2018-10-15 | 2021-02-04 | Air Force Medical University | Multi-target life detection method based on radar signal and detection radar |
-
2020
- 2020-08-27 CN CN202010880964.5A patent/CN112014821B/en active Active
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190317191A1 (en) * | 2018-04-11 | 2019-10-17 | Infineon Technologies Ag | Human Detection and Identification in a Setting Using Millimiter-Wave Radar |
US20190383904A1 (en) * | 2018-06-13 | 2019-12-19 | Metawave Corporation | Autoencoder assisted radar for target identification |
US20210033724A1 (en) * | 2018-10-15 | 2021-02-04 | Air Force Medical University | Multi-target life detection method based on radar signal and detection radar |
CN110007355A (en) * | 2019-04-15 | 2019-07-12 | 中国科学院电子学研究所 | The detection method and device of a kind of convolution self-encoding encoder and interior of articles exception |
CN110658507A (en) * | 2019-10-12 | 2020-01-07 | 电子科技大学 | Multi-class average maximization true and false target feature extraction method for radar target identification |
CN110691100A (en) * | 2019-10-28 | 2020-01-14 | 中国科学技术大学 | Hierarchical network attack identification and unknown attack detection method based on deep learning |
CN111144462A (en) * | 2019-12-16 | 2020-05-12 | 中国电子科技集团公司第三十八研究所 | Unknown individual identification method and device for radar signals |
CN111222575A (en) * | 2020-01-07 | 2020-06-02 | 北京联合大学 | KLXS multi-model fusion method and system based on HRRP target recognition |
CN111273288A (en) * | 2020-03-06 | 2020-06-12 | 电子科技大学 | Radar unknown target identification method based on long-term and short-term memory network |
CN111352086A (en) * | 2020-03-06 | 2020-06-30 | 电子科技大学 | Unknown target identification method based on deep convolutional neural network |
Non-Patent Citations (5)
Title |
---|
GUO, CHEN等: "Method for denoising and reconstructing radar HRRP using modified sparse auto-encoder", 《CHINESE JOURNAL OF AERONAUTICS》 * |
ZHU, ZHENGWEI等: "Automatic Target Recognition Based on HRRP Using SKO-KPCA", 《INTERNATIONAL JOINT CONFERENCE ON COMPUTATIONAL SCIENCES AND OPTIMIZATION, VOL 1, PROCEEDINGS》 * |
李斌等: "基于改进TSVM的未知网络应用识别算法", 《电子技术应用》 * |
田艺: "基于多样本与零样本学习的人体行为识别研究", 《中国优秀博硕士学位论文全文数据库(博士)信息科技辑》 * |
霍纬纲等: "基于自编码器和隐马尔可夫模型的时间序列异常检测方法", 《计算机应用》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113240011A (en) * | 2021-05-14 | 2021-08-10 | 烟台海颐软件股份有限公司 | Deep learning driven abnormity identification and repair method and intelligent system |
Also Published As
Publication number | Publication date |
---|---|
CN112014821B (en) | 2022-05-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109118479B (en) | Capsule network-based insulator defect identification and positioning device and method | |
CN111967294B (en) | Unsupervised domain self-adaptive pedestrian re-identification method | |
CN111181939B (en) | Network intrusion detection method and device based on ensemble learning | |
CN109389180A (en) | A power equipment image-recognizing method and inspection robot based on deep learning | |
CN108229550B (en) | Cloud picture classification method based on multi-granularity cascade forest network | |
CN108600135A (en) | A kind of recognition methods of signal modulation mode | |
CN107092829B (en) | Malicious code detection method based on image matching | |
CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
CN110147760B (en) | Novel efficient electric energy quality disturbance image feature extraction and identification method | |
CN108171119B (en) | SAR image change detection method based on residual error network | |
CN113435546B (en) | Migratable image recognition method and system based on differentiation confidence level | |
CN111524164B (en) | Target tracking method and device and electronic equipment | |
CN111343147B (en) | Network attack detection device and method based on deep learning | |
CN112766218B (en) | Cross-domain pedestrian re-recognition method and device based on asymmetric combined teaching network | |
CN108960142B (en) | Pedestrian re-identification method based on global feature loss function | |
CN116910752B (en) | Malicious code detection method based on big data | |
CN112801949A (en) | Method and device for determining discharge area in ultraviolet imaging detection technology | |
CN111310719B (en) | Unknown radiation source individual identification and detection method | |
CN112462355A (en) | Sea target intelligent detection method based on time-frequency three-feature extraction | |
CN112712102A (en) | Recognizer capable of simultaneously recognizing known radar radiation source individuals and unknown radar radiation source individuals | |
CN112014821B (en) | Unknown vehicle target identification method based on radar broadband characteristics | |
Du et al. | Convolutional neural network-based data anomaly detection considering class imbalance with limited data | |
CN113203954A (en) | Battery fault diagnosis method based on time-frequency image processing | |
CN109508350B (en) | Method and device for sampling data | |
CN116466408A (en) | Artificial neural network superbedrock identification method based on aeromagnetic data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |