Unmanned aerial vehicle and bird identification method based on confidence coding super-dimensional calculation

Abstract

The invention relates to an unmanned aerial vehicle and bird identification method based on confidence coding super-dimensional calculation. The method is mainly suitable for classification and identification of the unmanned aerial vehicle and the flying bird by a low-resolution radar with high repetition frequency detection capability. The method mainly comprises the following steps: in a training stage, firstly, preprocessing track information and micro-Doppler spectrum data, and constructing a fusion feature vector; estimating the confidence coefficient of each element of the feature vector, and carrying out vector weighted coding on the feature vector; constructing an unmanned aerial vehicle and bird training super-dimensional vector set; and in a test stage, performing similarity measurement on the test sample super-dimensional vector and the training super-dimensional vector set, and performing unmanned aerial vehicle and bird classification identification according to a calculation result. The method provided by the invention has the characteristics of simple engineering implementation, low system resource consumption, sufficient theoretical basis of the used method and the like, and can quickly and effectively solve the problem of identifying the unmanned aerial vehicle and the flying bird.

Description

Technical Field

[0001] The invention relates to a target recognition method for an unmanned aerial vehicle and a flying bird. Background Art

[0002] Drones and birds are typical "low, slow and small" targets with low observability. Effective monitoring and identification of the two has become a difficult problem that urgently needs to be solved to ensure air route safety and urban security. There are many types of drones and birds, low flight altitudes, strong maneuverability, and small radar cross-sectional areas. In addition, the detection environment is complex, which brings great trouble to the classification and identification of drones and birds. It has become a global problem. Therefore, it is urgent to develop "visible and identifiable" drones, birds and other "low, slow and small" target monitoring methods and technologies to achieve refined description and identification of targets.

[0003] At present, many drone and bird target identifications are based on track or micro-motion features. In radar measurement data, track is the most basic unit to characterize the target and is easy to obtain. The temporal and spatial information contained in the radar target track provides a basis for mining the differences in the behavioral patterns of drones and birds for target identification. For example, the feature extraction method based on target track proposed in the UAV radar track motion feature extraction and combined classification method in System Engineering and Electronic Technology in October 2023 uses a supervised learning method combined with a random forest model to distinguish the tracks of drones, birds, and precipitation clutter targets. Due to the high speed of the drone rotor, its micro-motion cycle is significantly faster than that of the bird, but the intensity is weaker than the micro-Doppler of the bird target. In addition, the irregular flapping of the wings caused by the bird maneuver makes its micro-Doppler more complex. There are obvious differences in the micro-motion features of drones and flying birds, which provide important feature basis for the accurate detection and precise identification of drone and flying bird targets that is independent of prior information and has good separability. For example, in 2017 IEEE Geoscience and Remote Sensing Letters, a convolutional neural network was used to realize drone target recognition based on micro-motion feature information. Summary of the invention

[0004] In order to solve the problem of low accuracy in the classification and identification of UAVs and flying birds by relying on a single feature in the prior art, the present invention proposes a method for classifying and identifying targets of UAVs and flying birds in a radar system.

[0005] The technical solution of the present invention includes:

[0006] First, the track information and micro-Doppler spectrum data are preprocessed to construct a fusion feature vector; then the confidence of each element of the feature vector is estimated, and the feature vector is vector-weighted encoded; a set of training hyperdimensional vectors for drones and birds is constructed; similarity is measured between the test sample hyperdimensional vector and the training hyperdimensional vector set, and drones and birds are classified and identified based on the calculation results. The specific steps include:

[0007] Step 1: Preprocess the track information and micro-Doppler spectrum data respectively to construct a fusion feature vector;

[0008] Step 2: Estimate the confidence of each element of the micro-Doppler spectrum part in the fused feature vector;

[0009] Step 3: Use the confidence estimation vector weighted coding method to encode the hyperdimensional vector of the training sample, mark the position of each element of the fused feature vector as an address, mark each element as the numerical value corresponding to the address, mark the confidence estimation result of each element as the weighted value of the numerical value corresponding to the address, calculate the address hyperdimensional vector and the numerical hyperdimensional vector respectively, and use the weighted values ​​of the address hyperdimensional vector, the numerical hyperdimensional vector and the numerical value corresponding to the address to encode the sample hyperdimensional vector;

[0010] Step 4: Encode each sample of the drone and the bird respectively to obtain the drone sample hyperdimensional vector and the bird sample hyperdimensional vector. Use the addition operation to group all the training hyperdimensional vectors belonging to the same category into a set.

[0011] Step 5: Retrain the training hyperdimensional vector set of the same category;

[0012] Step 6: Use the bubble difference distance method to measure the similarity between the test sample hyperdimensional vector and the training hyperdimensional vector set. Combined with the motion posture of the test sample, calculate the difference distance between the test sample and the training samples within the range of the drone and bird postures, and sort the difference distance set from small to large.

[0013] Step 7: Classify and identify drones and flying birds based on the calculation results;

[0014] Wherein the step 1 further comprises:

[0015] Step A: Extract the position, distance, elevation angle, heading, speed and amplitude information from the track;

[0016] Step B: Calculate the mean speed, speed variance, acceleration, amplitude mean, amplitude variance, and heading turning rate;

[0017] Step C: Shift the target spectrum in the micro-Doppler spectrum data to the zero frequency position;

[0018] Step D: normalizing the translated micro-Doppler spectrum;

[0019] Step E: Concatenate the preprocessed track and micro-Doppler spectrum to construct a feature vector;

[0020] Wherein said step 2 also comprises:

[0021] Step F: Counting the peak value and peak position of the micro-Doppler spectrum;

[0022] Step G: Calculate the signal-to-noise ratio of each peak value;

[0023] Step H: Calculate the confidence level of each frequency value of the micro-Doppler spectrum according to the peak value signal-to-noise ratio;

[0024] Wherein said step 5 also comprises:

[0025] Step I: Calculate the maximum correlation coefficient between each training hyperdimensional vector and the set;

[0026] Step J: Eliminate redundant training hyperdimensional vectors based on the maximum correlation coefficient.

[0027] The proposal and engineering implementation of the present invention have high promotion and application value in the field of drone and bird identification. It can effectively classify drone and bird targets under high-frequency radar detection of targets, and improve the accuracy of "low, slow and small" target recognition. The confidence estimation vector weighted coding method is used to encode the hyperdimensional vector of the training sample, which can simply and uniquely represent the training sample, with the characteristics of low system overhead, simple calculation, and easy engineering implementation. The bubble difference distance method is used to measure the similarity between the test sample hyperdimensional vector and the training hyperdimensional vector set, reducing the influence of background clutter and fixed clutter on track characteristics and micro-Doppler modulation.

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a data processing flow chart of a preferred embodiment of the present invention.

[0030] Figure 2 It is a schematic diagram of high repetition rate detection echo of UAV target according to a preferred embodiment of the present invention.

[0031] Figure 3 It is a schematic diagram of the micro-Doppler modulation characteristics of the UAV target according to the preferred embodiment of the present invention.

[0032] Figure 4 It is a schematic diagram of high repetition rate detection echo of flying bird targets according to a preferred embodiment of the present invention.

[0033] Figure 5 It is a schematic diagram of micro-Doppler modulation characteristics of a flying bird target in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention proposes a method for identifying UAVs and birds based on confidence coding hyperdimensional computing, which fully utilizes the track motion characteristics and micro-motion modulation characteristics of UAVs and birds, realizes the classification and identification of UAVs and birds through confidence coding hyperdimensional computing, and uses information more fully. Compared with deep learning methods such as convolutional neural networks, the present invention has low system resource occupancy, low training time overhead, and is easier to implement in engineering.

[0035] The preferred embodiment of the present invention is shown in the attached flow chart. Figure 1 , the steps include:

[0036] Step 1: Preprocess the track information and micro-Doppler spectrum data respectively to construct a fusion feature vector, including:

[0037] Step A, extracting the position, distance, elevation angle, heading, speed and amplitude information from the track;

[0038] Step B: Calculate the mean speed, speed variance, acceleration, amplitude mean, amplitude variance, and heading turning rate, specifically:

[0039] (1) Calculate the mean speed

[0040]

[0041] Where V i is the speed of the ith cycle, and N is the number of track cycles;

[0042] (2) Calculation of speed variance

[0043]

[0044] (3) Calculate acceleration

[0045]

[0046] Where T i+1 is the time of the i+1th cycle, T i is the time of the i-th cycle;

[0047] (4) Calculate the mean amplitude

[0048]

[0049] where F i is the amplitude of the ith cycle, and N is the number of track cycles;

[0050] (5) Calculate the amplitude variance

[0051]

[0052] (6) Calculate the heading turning rate

[0053]

[0054] in

[0055]

[0056] where h i is the heading of the ith cycle, h mean is the average heading, δ h is the heading deviation threshold.

[0057] Step C: Shift the target spectrum in the micro-Doppler spectrum data to the zero frequency position, specifically:

[0058] (1) Calculate the maximum value Fre of micro-Doppler spectrum data max and the location of the maximum value;

[0059] (2) With the maximum value as the center, the micro-Doppler spectrum is shifted until the maximum value moves to the zero frequency position.

[0060] Step D: normalize the shifted micro-Doppler spectrum, specifically:

[0061]

[0062] Among them, Fre min is the minimum value of the spectrum, and Fre is the micro-Doppler spectrum.

[0063] Step E: splice the preprocessed track and micro-Doppler spectrum to construct a feature vector, specifically:

[0064] The azimuth, distance, elevation angle, heading, heading turning rate, speed, speed mean, speed variance, acceleration, amplitude, amplitude mean, amplitude variance and normalized micro-Doppler spectrum are concatenated into a feature vector.

[0065] Step 2: Estimate the confidence of each element of the micro-Doppler spectrum in the fused feature vector, including:

[0066] Step F, counting the peak value and peak position of the micro-Doppler spectrum, specifically:

[0067] Traverse the micro-Doppler spectrum from the beginning. When Fre'(i)>Fre'(i-1) and Fre'(i)>Fre'(i+1), the i-th spectrum position is recorded as the peak position, and the spectrum value is recorded as the peak value Fre"(i)=Fre'(i).

[0068] Step G: Calculate the signal-to-noise ratio of each peak value, specifically:

[0069] Calculate the signal-to-noise ratio of each peak value based on the peak value and the noise mean

[0070]

[0071] Noise mean is the noise mean, obtained directly from the radar system.

[0072] Step H: Calculate the confidence of each frequency value of the micro-Doppler spectrum according to the peak value signal-to-noise ratio, specifically:

[0073] (1) Set the confidence level of each frequency value at the non-peak position in the micro-Doppler spectrum Fre' to 0;

[0074] (2) Further spectrum judgment is performed on the peak position. When the difference between two adjacent peak positions is less than 3, the peak position with a small signal-to-noise ratio is judged as a frequency sidelobe, and the confidence of the frequency value of this position is set to 0, and the confidence of the frequency values ​​of the other peak positions is set to.

[0075] Step 3: Encode the hyperdimensional vector of the training sample using the confidence estimation vector weighted encoding method, mark the position of each element of the fused feature vector as an address, mark each element as the numerical value corresponding to the address, mark the confidence estimation result of each element as the weighted value of the numerical value corresponding to the address, calculate the address hyperdimensional vector and the numerical hyperdimensional vector respectively, and use the weighted value of the address hyperdimensional vector, the numerical hyperdimensional vector and the numerical value corresponding to the address to encode the sample hyperdimensional vector, specifically:

[0076] (1) According to the number of addresses N a , select N completely randomly and without repetition from the d-dimensional hyperdimensional space a address hyperdimensional vector, denoted as

[0077] (2) Sort the values ​​corresponding to the addresses in ascending order;

[0078] (3) randomly selecting a hyperdimensional vector from the d-dimensional hyperdimensional space to represent the numerical hyperdimensional vector corresponding to the minimum value;

[0079] Then randomly flip a certain element in the previous numerical hyperdimensional vector to represent the hyperdimensional vector of the next data, and each element can only be flipped once;

[0080] (4) Repeat the previous step until all the values ​​corresponding to the addresses have completed the hyperdimensional vector encoding, and obtain N a A numerical hyperdimensional vector, denoted by

[0081] (5) The address hyperdimensional vector corresponding to each address in the sample, the data hyperdimensional vector corresponding to the value in each address, and the weighted value of the value corresponding to the address are bound into an address-weighted value pair through multiplication operation, and then all address-weighted value pairs are encoded into the hyperdimensional vector of the sample through addition operation.

[0082]

[0083] in is the multiplication operator for hyperdimensional vectors.

[0084] Step 4: Encode each sample of the drone and the bird respectively to obtain the drone sample hyperdimensional vector and the bird sample hyperdimensional vector. Use the addition operation to group all the training hyperdimensional vectors belonging to the same category into a set, specifically:

[0085] Each sample of the drone and the bird is weighted encoded based on the confidence estimation vector to obtain N F The hyperdimensional vector of drone samples and N M The hyperdimensional vectors of the flying bird samples are added again to form a set of all the training hyperdimensional vectors belonging to the same category, which are recorded as C1 and C2 respectively.

[0086]

[0087] Step 5: Retrain the training hyperdimensional vector set of the same category, including:

[0088] Step I: Calculate the maximum correlation coefficient between each training hyperdimensional vector and the set, specifically:

[0089]

[0090] Where R i ,R j are the i-th and j-th training hyperdimensional vectors in the class, Cov(R i ,R j ) is the covariance of the i-th and j-th training hyperdimensional vectors, D(R i ), D(R j ) are the variances of the i-th and j-th training hyperdimensional vectors, respectively.

[0091] Step J: Eliminate redundant training hyperdimensional vectors according to the maximum correlation coefficient, specifically:

[0092] When i,j >0.7, the i-th and j-th training hyperdimensional vector samples are considered close, and any one of them is removed, and all training hyperdimensional vectors belonging to the same category are reorganized into a set, which are recorded as C1' and C2' respectively.

[0093]

[0094] Where N F ' is the number of hyperdimensional vectors of drone samples after removing redundancy, N M ' is the number of hyperdimensional vectors of the bird samples after removing redundancy.

[0095] Step 6: Use the bubble difference distance method to measure the similarity between the test sample hyperdimensional vector and the training hyperdimensional vector set. Combined with the motion posture of the test sample, calculate the difference distance between the test sample and the training sample within the range of the drone and bird posture, and sort the difference distance set from small to large. The method is as follows:

[0096] (1) UAV and bird training samples within the ±15 degree range of the test sample posture are selected as the basic samples for similarity measurement;

[0097] (2) Constructing the outlier difference distance function

[0098]

[0099] Where R Test is the test sample hyperdimensional vector, C j ' is the jth training hyperdimensional vector in the basic sample, d(R Test ,C j ')=||R Test ,C j '|| represents the Euclidean distance between two hyperdimensional vectors, and the parameter β is the mean of the Euclidean distance between the hyperdimensional vectors of the basic samples, which is used to control D(R Test ,C j ')'s growth rate;

[0100] (3) Calculate the outlier difference distances between the test sample hyperdimensional vector and the drone and bird base samples respectively, and form the outlier difference distance and label set {{D i},{Lable i}};

[0101] (4) For the outlier difference distance set {D i The elements in} are sorted from small to large according to their numerical values ​​to form a similarity measurement set {{D j},{Lable j}}, where {D j} is the sorted set, {Lable j} is the corresponding label after sorting.

[0102] Step 7: Classify and identify drones and birds based on the calculation results, specifically:

[0103] According to the sorted set {Dj}, take the first value as the recognition confidence, and take the first corresponding label as the classification result.

Claims

1. A method for identifying drones and flying birds based on confidence-coded hyperdimensional computing, characterized in that: Step 1: Preprocess the track information and micro-Doppler spectrum data respectively to construct a fusion feature vector; Step 2: Estimate the confidence of each element of the micro-Doppler spectrum part in the fused feature vector; Step 3: Use the confidence estimation vector weighted coding method to encode the hyperdimensional vector of the training sample, mark the position of each element of the fused feature vector as an address, mark each element as the numerical value corresponding to the address, mark the confidence estimation result of each element as the weighted value of the numerical value corresponding to the address, calculate the address hyperdimensional vector and the numerical hyperdimensional vector respectively, and use the weighted values ​​of the address hyperdimensional vector, the numerical hyperdimensional vector and the numerical value corresponding to the address to encode the sample hyperdimensional vector; Step 4: Encode each sample of the drone and the bird respectively to obtain the drone sample hyperdimensional vector and the bird sample hyperdimensional vector. Use the addition operation to group all the training hyperdimensional vectors belonging to the same category into a set. Step 5: Retrain the training hyperdimensional vector set of the same category; Step 6: Use the bubble difference distance method to measure the similarity between the test sample hyperdimensional vector and the training hyperdimensional vector set. Combined with the motion posture of the test sample, calculate the difference distance between the test sample and the training samples within the range of the drone and bird postures, and sort the difference distance set from small to large. Step 7: Classify and identify drones and flying birds based on the calculation results; Wherein the step 1 further comprises: Step A: Extract the position, distance, elevation angle, heading, speed and amplitude information from the track; Step B: Calculate the mean speed, speed variance, acceleration, amplitude mean, amplitude variance, and heading turning rate; Step C: Shift the target spectrum in the micro-Doppler spectrum data to the zero frequency position; Step D: normalizing the translated micro-Doppler spectrum; Step E: Concatenate the preprocessed track and micro-Doppler spectrum to construct a feature vector; Wherein said step 2 further comprises: Step F: Counting the peak value and peak position of the micro-Doppler spectrum; Step G: Calculate the signal-to-noise ratio of each peak value; Step H: Calculate the confidence level of each frequency value of the micro-Doppler spectrum according to the peak value signal-to-noise ratio; Wherein said step 5 also comprises: Step I: Calculate the maximum correlation coefficient between each training hyperdimensional vector and the set; Step J: Eliminate redundant training hyperdimensional vectors based on the maximum correlation coefficient.

2. The method for identifying drones and flying birds based on confidence-coded hyperdimensional computing according to claim 1, characterized in that: The step 3 uses the confidence estimation vector weighted coding method to encode the super-dimensional vector of the training sample, specifically: first, according to the number of addresses N a , select N completely randomly and without repetition from the d-dimensional hyperdimensional space a address hyperdimensional vector, denoted as Secondly, sort the values ​​corresponding to the addresses in ascending order; A hyperdimensional vector is randomly selected from the d-dimensional hyperdimensional space to represent the numerical hyperdimensional vector corresponding to the minimum value; Then randomly flip a certain element in the previous numerical hyperdimensional vector to represent the hyperdimensional vector of the next data, and each element can only be flipped once; Repeat the previous step until the values ​​corresponding to the addresses have completed the hyperdimensional vector encoding, and get N a A numerical hyperdimensional vector, denoted by Finally, the address hyperdimensional vector corresponding to each address in the sample, the data hyperdimensional vector corresponding to the value in each address, and the weighted value of the value corresponding to the address are bound into address-weighted value pairs through multiplication operations, and then all address-weighted value pairs are encoded into the hyperdimensional vector of the sample through addition operations: in is the multiplication operator for hyperdimensional vectors.

3. The method for identifying drones and flying birds based on confidence-coded hyperdimensional computing according to claim 1, characterized in that: The step 6 uses the bubble difference distance method to measure the similarity between the test sample hyperdimensional vector and the training hyperdimensional vector set, specifically: Firstly, the UAV and bird training samples within the range of ±15 degrees of the test sample posture are selected as the basic samples for similarity measurement; Next, construct the outlier difference distance function: Where R Test is the test sample hyperdimensional vector, C j is the jth training hyperdimensional vector in the basic sample, d(R Test ,C j )=||R Test ,C j || represents the Euclidean distance between two hyperdimensional vectors, and the parameter β is the mean of the Euclidean distance between the basic sample hyperdimensional vectors, which is used to control D(R Test ,C j )’s growth rate; Then, the outlier difference distances between the test sample hyperdimensional vector and the drone and bird base samples are calculated to form the outlier difference distance and label set {{D i },{Lable i }}; Finally, the outlier difference distance set {D i The elements in} are sorted from small to large according to their numerical values ​​to form a similarity measurement set {{D j },{Lable j }}, where {D j } is the sorted set, {Lable j } is the corresponding label after sorting.

Images