Underwater slow small target classification method using trajectory features and joint classifier

Abstract

The present invention proposes a classification method for underwater slow small targets using trajectory features and joint classifiers. The feature quantity is calculated by feature extraction of the trajectory, and then the feature quantity is input into the pre-trained SVDD-SVM joint classifier. Finally, the classification result is obtained by comprehensive evaluation of the joint classifier output using voting criteria. The proposed classification method based on the joint classifier has a recall rate and precision rate that are superior to those of the traditional classifier, and can solve the classification problems caused by small samples and class imbalance to a certain extent. The underwater slow small target classification method extracted by the present invention using tracking trajectory features and joint classifiers has a higher recall rate and precision rate: the average recall rate for frogman targets can reach 86%, and the average precision rate can reach 87%. The average recall rate of SVDD-SVM for UUV targets can reach 85%, and the average precision rate can reach 86%, which can alleviate the situation where the classification performance of underwater slow small targets is reduced due to unstable features, small sample data, and class imbalance.

Description

Technical Field

[0001] The invention belongs to the field of underwater recognition, and in particular relates to a method for classifying underwater slow small targets using trajectory features and a joint classifier. Background Art

[0002] Underwater slow small targets include frogmen, unmanned underwater vehicles (UUVs), autonomous underwater vehicles (AUVs), and fish. Most of the current research targets are frogmen and UUVs.

[0003] Traditional underwater slow small target classification and recognition methods use support vector data description (SVDD), support vector machine (SVM), neural network (NN) and other machine learning methods to classify underwater slow small targets based on target morphological characteristics and radiation noise characteristics. However, underwater slow small targets have unclear morphological characteristics and weak radiation noise. In addition, their data has the characteristics of small samples and class imbalance, which leads to limited performance of these traditional classification methods. Although the classification performance of these traditional methods can be improved through methods such as transfer learning and energy statistics, they still cannot fundamentally overcome their inherent defects of being unsuitable for small samples and class imbalance data.

[0004] Unlike morphological features and radiation noise features, tracking trajectory features are less affected by the degree of small target morphological features and radiation noise intensity, and they also have the potential to classify slow small targets. However, the classification method using tracking trajectory features still faces the problem of small sample data and class imbalance. It needs to be combined with a classification method with small sample and class imbalance data classification capabilities to give full play to its advantages in small target classification. Summary of the invention

[0005] The technical problem solved by the present invention is: in order to solve the problem of reduced classification performance caused by small sample data and class imbalance faced by underwater slow small target classification, the present invention proposes an underwater slow small target classification method using trajectory features and a joint classifier.

[0006] The technical solution of the present invention is: a method for classifying underwater slow small targets using trajectory features and a joint classifier, comprising the following steps:

[0007] Step 1: Extract the tracking trajectory feature of underwater slow small target;

[0008] Step 2: Design SVDD-SVM joint classifier: This joint classifier adopts a two-level classifier series architecture; the first-level classifier contains two parallel SVDDs, and the second-level classifier contains a trained SVM, which is connected in series after the two SVDDs;

[0009] Step 3: Train and test the SVDD-SVM joint classifier;

[0010] Step 4: Design the voting criteria required for classification for the trained and tested SVDD-SVM joint classifier: If the two SVDDs obtain mutually supportive results, then the result is directly used as the final classification result; if the two SVDDs obtain contradictory results, then the result of the SVM classifier is used as the final classification result;

[0011] Step 5: For the tracking trajectory to be classified, extract the feature quantity according to step 1, and use the SVDD-SVM joint classifier trained and tested in step 3 to classify the tracking trajectory according to the voting criteria in step 4, so as to achieve the classification of underwater slow small targets.

[0012] Furthermore, in step 1, for the local feature quantity of the underwater slow small target, the minimum value, maximum value, mean value, and variance are extracted as feature quantities; for the global feature quantity of the underwater slow small target, it can be directly used as the feature quantity, and the sum of the two types of feature quantities is used as the underwater slow small target tracking trajectory feature quantity.

[0013] Furthermore, in step 3, the tracking trajectory is used as training data, the feature quantity is extracted according to step 1 to construct a training set, and the SVDD-SVM joint classifier is trained, and the SVDD and SVM are trained separately.

[0014] Furthermore, when training SVDD, the frogman tracking trajectory is used as training to extract feature quantities, construct a training set, and train SVDD No. 1, and its output results are "frogman" and "others"; the UUV tracking trajectory is used as training to extract feature quantities, construct a training set, and train SVDD No. 2, and its output results are "UUV" and "others".

[0015] Furthermore, when training the SVM, the SVM can classify the confused data in the confused region of the SVDD classifier, and the output results are “frogman” and “UUV”.

[0016] Furthermore, in step 3, the tracking trajectory is used as the test data, and the feature quantity is extracted according to step 1 to construct a test set to test the SVDD-SVM joint classifier; during the test, the final classification output result is given by the output results of the two SVDDs and SVM and the voting criteria.

[0017] Furthermore, in step 4, the voting criteria specifically include the following steps:

[0018] Step 4.1: Send the test data to the two first-level SVDDs synchronously. When the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "Others", classify the data as "Frogman"; when the output of SVDD No. 1 is "Others" and the output of SVDD No. 2 is "UUV", classify the data as "UUV"; when the outputs of SVDD No. 1 and SVDD No. 2 are both "Others", classify the data as "Others"; when the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "UUV", send the contradictory test data to the second-level SVM;

[0019] Step 4.2: Take the output of the second-level single SVM as the output of the joint classifier: when the SVM output is “frogman”, the data is classified as “frogman”; when the SVM output is “UUV”, the data is classified as “UUV”.

[0020] Effects of the Invention

[0021] The technical effect of the present invention is that the present invention calculates feature quantities by extracting features from trajectories, then inputs the feature quantities into a pre-trained SVDD-SVM joint classifier, and finally uses voting criteria to comprehensively evaluate the output of the joint classifier to obtain a classification result. The proposed classification method based on the joint classifier has a better recall rate and precision rate than the traditional classifier, and can solve the classification problems caused by small samples and class imbalance to a certain extent.

[0022] The basic principle and implementation scheme of the present invention have been verified by measured data and Monte Carlo experiments, and the results show that the classification method of underwater slow small targets using tracking trajectory features and joint classifiers has higher recall rate and precision rate: the average recall rate of frogman targets can reach 86%, and the average precision rate can reach 87%. The average recall rate of SVDD-SVM for UUV targets can reach 85%, and the average precision rate can reach 86%, which to a certain extent solves the problem of reduced classification performance of underwater slow small targets due to unstable features, small sample data, and class imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the basic architecture and information flow of the SVDD-SVM joint classifier;

[0024] Figure 2 Schematic diagram of the training process of two SVDD classifiers;

[0025] Figure 3 Schematic diagram of the training process for a single SVM classifier;

[0026] Figure 4 The main steps of the present invention are as follows;

[0027] FIG5 is a schematic diagram of a small target detection test, wherein FIG5(a) is a schematic diagram of frogman detection, and FIG5(b) is a schematic diagram of UUV detection;

[0028] Figure 6 shows the detection results of small target detection sonar, where Figure 6(a) is a single frame detection output and Figure 6(b) is a multi-target tracking output;

[0029] Figure 7 is the confusion matrix of SVM, SVDD and SVDD-SVM results (where the yellow area indicates that the predicted category is the same as the true category, and the white area indicates that the predicted category is different from the true category), where Figure 7(a) is the confusion matrix of the SVM classification result, where Figure 7(b) is the confusion matrix of the SVDD classification result, and where Figure 7(c) is the confusion matrix of the SVDD-SVM classification result;

[0030] Figure 8 shows the results of 100 Monte Carlo repeated experiments, where Figure 8(a) is the frogman recall rate, Figure 8(b) is the frogman precision rate, Figure 8(c) is the UUV recall rate, Figure 8(d) is the UUV precision rate, and Figure 8(e) is the accuracy rate;

[0031] Table 1 shows the tracking trajectory characteristics of slow small underwater targets;

[0032] Table 2 shows the voting rules for the SVDD-SVM classifier;

[0033] Table 3 shows the average classification performance comparison of SVM, SVDD and SVDD-SVM. DETAILED DESCRIPTION

[0034] See also Figure 1 - Figure 8, the main contents of this method are:

[0035] (1) The tracking trajectory feature of underwater slow small target is designed. The mean, variance, maximum value and minimum value of three features, namely motion speed, trajectory smoothness and direction change rate, are extracted, and a total of 12 feature quantities are used as local feature quantities. The trajectory curvature, the distance between the starting and ending points, and the distance between trajectory points are extracted as global feature quantities. The local feature quantities and the global feature quantities are connected in series to form a 15-dimensional underwater slow small target tracking trajectory feature quantity.

[0036] (2) To address the problem of small samples and class imbalance, an SVDD-SVM joint classifier was designed based on trajectory features. The SVDD-SVM joint classifier was designed using a two-stage classifier series architecture. Two parallel single-classification SVDDs were used as the first-stage classifier, and a binary classification SVM was connected in series after the SVDD as the second-stage classifier. The information flow of the SVDD-SVM joint classifier during classification is as follows: first, the tracking trajectory features are sent to the two SVDDs of the first stage respectively, and the subsequent information flow is divided into two types according to the voting criteria. One is to use the classification result of the SVDD as the final result of the SVDD-SVM joint classifier; the other is to let the tracking trajectory features continue to be input into the second-stage SVM, and the classification result of the SVM is used as the final result of the SVDD-SVM joint classifier. According to the above overall architecture and information flow design, the SVDD-SVM joint classifier training process that matches the following is implemented: the first-level SVDD No. 1 is trained using the frogman’s tracking trajectory feature quantity, and the output results are “frogman” and “others”; the SVDD No. 2 is trained using the UUV’s tracking trajectory feature quantity, and the output results are “UUV” and “others”; the SVM is trained using the same frogman and UUV’s tracking trajectory feature quantities as those used by SVDD No. 1 and SVDD No. 2, and the output results are “frogman” and “UUV”.

[0037] (3) The voting criteria of the SVDD-SVM joint classifier are designed. The tracking trajectory features are synchronously sent to the two SVDDs of the SVDD-SVM joint classifier. If the output results of the two SVDDs support each other, that is, when the output of SVDD No. 1 is "frogman" and the output of SVDD No. 2 is "other", the data is classified as "frogman"; when the output of SVDD No. 1 is "other" and the output of SVDD No. 2 is "UUV", the data is classified as "UUV"; when the outputs of SVDD No. 1 and SVDD No. 2 are both "other", the data is classified as "other". If the outputs of the two SVDDs are contradictory, that is, the output of SVDD No. 1 is "frogman" and the output of SVDD No. 2 is "UUV", the contradictory test data is sent to the second-level SVM. When the SVM output is "frogman", the data is classified as "frogman"; when the SVM output is "UUV", the data is classified as "UUV".

[0038] (4) The processing results of the underwater slow small target classification method using tracking trajectory features and joint classifiers are given through the measured tracking data of underwater slow small targets. This proves that the proposed classification method based on the joint classifier has better recall rate and precision than the traditional classifier, and can solve the classification problems caused by small samples and class imbalance to a certain extent.

[0039] In this embodiment, the underwater slow small targets studied are frogmen and UUVs, and the classification and identification of frogmen and UUVs are completed through the following four steps.

[0040] In general, this method consists of the following five steps:

[0041] Step 1: Extract the tracking trajectory features of slow small underwater targets.

[0042] Frogmen and UUVs are affected by the underwater environment to different degrees, and their own power sources and navigation methods are different, which leads to a certain degree of difference between the trajectories of frogmen and UUVs. These differences are quantitatively analyzed, and in this embodiment, four aspects are studied: movement speed, trajectory smoothness, trajectory length, and trajectory directionality.

[0043] (1) Movement speed

[0044] Due to physical limitations, frogmen generally swim at a low speed underwater, between 1 and 2 knots. The speed of the UUV varies with the model, working conditions, and purpose of use, but in many cases its speed is greater than that of the frogmen, and can reach more than 2 knots. At the same time, frogmen are affected by their own physical strength during movement, and their movement speed shows the characteristics of "fast first and then slow". When the trajectory lasts for a long time, the frogman's movement speed will change significantly. Without considering the direction, the movement speed is calculated using the coordinate points p i The velocity magnitude v i express:

[0045]

[0046] In the formula, v i is the velocity of the ith point, p i =(x i ,y i ) T is the coordinate of the i-th (i=1,2,…,N) point, x i is the horizontal coordinate of the i-th point, y i is the ordinate of the i-th point, || ||2 represents the Euclidean norm distance, and ΔT is the interval time between two adjacent measurements during the tracking process.

[0047] (2) Trajectory smoothness

[0048] The frogman is easily affected by the disturbance of water flow and the uncertainty of its own motion, so its trajectory is prone to "burrs" and poor smoothness. In contrast, the UUV follows a pre-set path, so its trajectory has fewer "burrs" and good smoothness. The trajectory smoothness is calculated using the trajectory point p i To the line connecting the starting and ending points of the trajectory (i.e. passing through p1 and p at the same time N The offset distance D of the straight line iThe offset distance D i The smaller it is, the better the trajectory smoothness. i It can be expressed as:

[0049]

[0050] Where D i is the offset distance of the i-th point, p1=(x1,y1) T is the coordinate of the starting point of the trajectory, p N =(x N ,y N ) T is the coordinate of the end point of the trajectory, x i is the horizontal coordinate of the i-th point, y i is the ordinate of the i-th point.

[0051] (3) Track length

[0052] Due to the limitation of physical factors, frogmen cannot stay in motion all the time and are prone to stop or nearly stop. Therefore, there are often stationary points in the frogmen's trajectory, resulting in a short continuous motion trajectory. UUVs are propulsed by propellers and rarely hover in the water. There are few stationary points in their trajectories, so the continuous motion trajectory is longer. The trajectory length uses the distance between the starting and ending points L and the sum of the distances between the trajectory points R as feature quantities, which are:

[0053]

[0054]

[0055] Where L is the distance between the starting and ending points, R is the sum of the distances between the trajectory points, and p i =(x i ,y i ) T is the coordinate of the i-th (i=1,2,…,N) point, x i is the horizontal coordinate of the i-th point, y i is the ordinate of the i-th point, and || ||2 represents the Euclidean norm distance.

[0056] (4) Track direction

[0057] Due to the low visibility of shallow water environment, frogmen lack direction guidance when swimming underwater, and are prone to deviate from the original established route, which needs to be corrected continuously, and the direction of their trajectory is unstable. UUV adopts an autonomous navigation mechanism, and its trajectory direction is more stable. The trajectory direction uses curvature ρ and direction change rate s i As feature quantity. Curvature ρ and direction change rate s iThe larger the curvature, the more times the target changes direction during the duration of the trajectory, and vice versa. The curvature ρ represents the ratio of the distance between the trajectory points and the distance between the starting and ending points of the trajectory:

[0058]

[0059] Where ρ is the curvature, L is the distance between the start and end points, and R is the sum of the distances between trajectory points.

[0060] Direction change rate i It represents the ratio of the sum of the two distances of a set of three adjacent coordinate points to the distance of the starting point among the three points:

[0061]

[0062] In the formula, s i is the rate of change of direction, p i =(x i ,y i ) T is the coordinate of the i-th (i=1,2,…,N) point, x i is the horizontal coordinate of the i-th point, y i is the ordinate of the i-th point, and ||||2 is the Euclidean norm distance.

[0063] In the above characteristics, the movement speed v i , distance D i , Direction change rate s i It is not a constant value during the duration of the trajectory and is called the local feature of the trajectory. For these three local features, since their changes also reflect the differences between different target trajectories, their respective minimum, maximum, mean, and variance are used as feature quantities. A total of 12 feature quantities are obtained from the three local features. On the contrary, the distance L between the start and end points, the distance and R between the trajectory points, and the curvature ρ do not change during the duration of the trajectory and are called the global feature of the trajectory. A total of 3 feature quantities are obtained from the three global features. In summary, there are a total of 15 feature quantities used, namely 12 local feature quantities and 3 global feature quantities, which are summarized in Table 1.

[0064] After obtaining all the above 15 feature quantities, the features need to be normalized. In this embodiment, min-max normalization is used. Suppose the feature vector corresponding to each feature quantity is f l (l=1,2,…,15), the new feature quantity after normalization is The relationship between the two can be expressed as:

[0065]

[0066] In the formula, min() means taking the minimum value, max() means taking the maximum value, and fl Represents the th feature quantity.

[0067] Step 2: Design an SVDD-SVM joint classifier suitable for underwater slow small target small sample and class imbalance data classification. The SVDD-SVM joint classifier consists of two classifiers, SVDD and SVM, which are introduced below.

[0068] Principles of the two classifiers.

[0069] (1) Support data description (SVDD)

[0070] SVDD is a single-classification algorithm whose goal is to find a hypersphere as small as possible in the feature space so that it can include as many target class samples as possible, thereby distinguishing between target samples and non-target samples. SVDD is usually used in the field of abnormal data detection.

[0071] Assume that the target class training sample is a column vector x i (i=1,2,…,N), the problem of finding a hypersphere with the smallest volume can be transformed into the following convex optimization problem:

[0072]

[0073] Where min() means taking the minimum value, r is the center of the hypersphere, C = 1 / (Nν) is the regularization parameter, N is the number of samples, ν∈(0,1] is the rejection rate of abnormal target class samples, indicating that a certain training error is allowed (that is, some target samples are allowed not to fall into the hypersphere), ξ i is the i-th relaxation factor, δ is the center of the hypersphere. C is used to balance the volume of the hypersphere and the error. Relaxation factor ξ i And the regularization parameter C prevents overfitting during training.

[0074] The Lagrange multiplier method is introduced to transform the original problem of equation (8) into a dual problem:

[0075]

[0076] In the formula, min() means taking the minimum value, x i is the i-th target class training sample, α i For sample x i The corresponding Lagrangian coefficient, α j (Subscript j and subscript i have the same definition) is the sample x j The corresponding Lagrangian coefficient, T represents transposition, C = 1 / (Nν) is the regularization parameter, N is the number of samples, and ν∈(0,1] refers to the rejection rate of abnormal target class samples. Satisfying 0<a iThe samples corresponding to the Lagrange coefficients of <C are called support vectors. The set of samples belonging to the support vectors in the training set is set as SV.

[0077] The data will not be completely spherically distributed in the feature space, and a kernel function needs to be introduced to map it to a high-dimensional feature space. The optimization problem corresponding to SVDD can be further rewritten as:

[0078]

[0079]

[0080] In the formula, min() means taking the minimum value, ||||2 is the Euclidean norm distance, K(·) represents the inner product operation after mapping to the high-dimensional feature space, and x i is the i-th target class training sample, α i For sample x i The corresponding Lagrangian coefficient, α j (Subscript j and subscript i have the same definition) is the sample x j The corresponding Lagrangian coefficient, C = 1 / (Nν) is the regularization parameter, N is the number of samples, and ν∈(0,1] refers to the rejection rate of abnormal target class samples.

[0081] (2) Support Vector Machine (SVM)

[0082] SVM is a classifier that minimizes structural risk. SVM finds a hyperplane in the feature sample space that can divide the target into two categories. If the feature distribution of the sample is complex, it is difficult to find a hyperplane that can completely divide the sample correctly. This can be achieved by introducing a relaxation factor ξ i With the regularization parameter C, the SVM problem is transformed into a convex optimization problem:

[0083]

[0084] Where min() means taking the minimum value, || ||2 is the Euclidean norm distance, w is the normal vector of the hyperplane with solution, C = 1 / (Nν) is the regularization parameter, N is the number of samples, ν∈(0,1] refers to the rejection rate of abnormal target class samples, ξ i is the ith relaxation factor, c i is the i-th target category, x i is the i-th target class training sample, and b is the offset.

[0085] Introducing the Lagrange multiplier α i , μ i , transforming equation (11) from an inequality constraint to an equality constraint, and obtaining the dual problem expression of equation (11):

[0086]

[0087] In the formula, max() means taking the maximum value, α i is the i-th Lagrange multiplier coefficient of the first constraint, c i is the i-th target category, x i is the i-th target class training sample, T represents transposition, C = 1 / (Nν) is the regularization parameter, N is the number of samples, ν∈(0,1] refers to the rejection rate of abnormal target class samples, μ i is the i-th Lagrange multiplication coefficient of the second constraint.

[0088] Similar to SVDD, SVM needs to use kernel functions to map sample features to a higher-dimensional feature space and divide hyperplanes in the high-dimensional feature space to achieve sample classification.

[0089] The SVM optimization problem after adding the kernel function can be expressed as:

[0090]

[0091] In the formula, min() means taking the minimum value, α i is the i-th Lagrange multiplier coefficient of the first constraint, c i is the i-th target category, K(·) represents the inner product operation after mapping to the high-dimensional feature space, C = 1 / (Nν) is the regularization parameter, N is the number of samples, ν∈(0,1] refers to the rejection rate of abnormal target class samples, μ i is the i-th Lagrange multiplication coefficient of the second constraint.

[0092] (3) SVDD-SVM joint classifier

[0093] After constructing two single-classification SVDDs for frogmen and UUVs respectively using training samples, a two-class classifier that can target frogmen and UUVs at the same time is constructed. When the two SVDDs are confused with frogmen and UUVs (the output results of the two SVDDs are frogmen and UUVs respectively, which lead to contradictions), the classifier can be used to further classify the confused samples. Since the newly constructed classifier needs to be able to further distinguish from the frogman and UUV confusion areas, a stronger classifier can be used. Considering that the two single-classification SVDDs are designed for frogmen and UUV data respectively, the two-classification SVM for frogmen and UUVs is used here to connect in series with the two single-classification SVDDs to form a joint classifier. The expression of the single-classification SVDD is shown in formula (10). The expression of the two-classification SVM is shown in formula (13).

[0094] The combined classifier of SVDD and SVM is called SVDD-SVM. Its basic architecture and information flow diagram are shown in the following figure. Figure 1 The joint classifier takes into account the characteristic distribution of frogmen and UUV samples (including the distribution of frogmen and UUV samples in the hypersphere), and draws on the idea of ​​improving strong classifiers by multiple weak classifiers in ensemble learning, which can alleviate the confusion between frogmen and UUV characteristic samples.

[0095] Depend on Figure 1 It can be seen that the SVDD-SVM joint classifier adopts a two-level classifier series architecture. Specifically, the first-level classifier contains two trained parallel SVDDs, and the second-level classifier contains one trained SVM, which is connected in series after the two SVDDs.

[0096] Step 3: Use the tracking trajectory as training data and extract the features according to step 1) to construct a training set.

[0097] Training the SVDD-SVM joint classifier. Using the tracking trajectory as the test data, extract the feature quantity according to step 1) to construct a test set, and test the SVDD-SVM joint classifier.

[0098] During the training process, the SVDD and SVM of the SVDD-SVM joint classifier are trained separately. If this training process is the first time, all the parameters of the two SVDDs and one SVM are initialized to 1; if this training is not the first time, the parameters adjusted in the last step are used as all the parameters of the two SVDDs and one SVM. When training the SVDD, the frogman tracking trajectory used as training is used to extract feature quantities, and a training set is constructed as the input of SVDD No. 1, and its output results are "frogman" and "others". According to the output results, the training accuracy of SVDD No. 1 is obtained by dividing the number of correctly classified samples by the total number of samples in the training set. The feature quantities are extracted using the UUV tracking trajectory used as training, and a training set is constructed as the input of SVDD No. 2, and its output results are "UUV" and "others". According to the output results, the training accuracy of SVDD No. 2 is obtained by dividing the number of correctly classified samples by the total number of samples in the training set. The schematic diagram of the training process of the two SVDD classifiers is shown in the figure. Figure 2 shown.

[0099] The same frogman and UUV training set is used as the input of SVM. The SVM can classify the confused data in the confused area of ​​SVDD classifier, and the output result is "frogman" and "UUV". According to the output result, the number of correctly classified samples is divided by the total number of samples in the training set to get the SVM training accuracy. The training process of a single SVM classifier is shown in the figure below. Figure 3 shown.

[0100] After training, the frogman and UUV trajectories used as tests were used to extract feature quantities, construct a test set, and test the SVDD-SVM joint classifier.

[0101] When testing SVDD, the frogman tracking trajectory is used as a test to extract feature quantities and construct a test set as the input of SVDD No. 1, and its output results are "frogman" and "others". According to the output results, the number of correctly classified samples is divided by the total number of samples in the test set to obtain the test accuracy of SVDD No. 1. The UUV tracking trajectory is used as a test to extract feature quantities and construct a test set as the input of SVDD No. 2, and its output results are "UUV" and "others". According to the output results, the number of correctly classified samples is divided by the total number of samples in the test set to obtain the test accuracy of SVDD No. 2.

[0102] Using the same frogman and UUV test set as SVM input, the output results are "frogman" and "UUV". Based on the output results, the SVM test accuracy is obtained by dividing the number of correctly classified samples by the total number of samples in the test set.

[0103] Calculate the difference between the training accuracy of SVDD No. 1 and the test accuracy of SVDD No. 1, the difference between the training accuracy of SVDD No. 2 and the test accuracy of SVDD No. 2, and the difference between the training accuracy of SVM and the test accuracy of SVM. If all three are less than 5%, proceed to step 4. Otherwise, between 0.0001 and 10 10 Adjust the parameters of SVDD No. 1, SVDD No. 2, and SVM and repeat step 3.

[0104] Step 4: Design the voting criteria required for classification for the trained and tested SVDD-SVM joint classifier.

[0105] For different combinations of the output results of two SVDDs and SVMs, the voting criteria of the designed SVDD-SVM joint classifier are shown in Table 2. The voting criteria in Table 2 can be summarized as follows: if the two SVDDs obtain mutually supportive results, then the result is directly used as the final classification result; if the two SVDDs obtain contradictory results, then the result of the SVM classifier is used as the final classification result. Figure 1 , Figure 2 , Figure 3 The architecture of the SVDD-SVM joint classifier voting criterion can be described in detail as:

[0106] In the first step, the test data is synchronously sent to the two first-level SVDDs. If the outputs of the two SVDDs support each other, that is, when the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "Other", the data is classified as "Frogman"; when the output of SVDD No. 1 is "Other" and the output of SVDD No. 2 is "UUV", the data is classified as "UUV"; when the outputs of SVDD No. 1 and SVDD No. 2 are both "Other", the data is classified as "Other". If the outputs of the two SVDDs are contradictory, that is, the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "UUV", the contradictory test data is sent to the second-level SVM.

[0107] In the second step, the output of the second-level single SVM is used as the output of the joint classifier. When the SVM output is "frogman", the data is classified as "frogman"; when the SVM output is "UUV", the data is classified as "UUV".

[0108] Step 5: Classify the tracking trajectories to be classified, so as to achieve the classification of slow small underwater targets.

[0109] The tracking trajectory to be classified is used to extract the feature quantity according to step 1, and the trajectory feature quantity is used as the input of the SVDD-SVM joint classifier trained and tested in step 3. The tracking trajectory is classified according to the voting criteria in step 4, thereby realizing the classification of underwater slow small targets.

[0110] The content of this method is further explained below based on a specific example.

[0111] Taking the typical classification of slow underwater small targets as an example, an implementation example of the present invention is given. The implementation example verifies that the classification method of slow underwater small targets using tracking trajectory features and joint classifiers has higher precision, recall and accuracy based on the classification results of measured data.

[0112] Brief introduction of measured data

[0113] The wet end of the small target detection sonar is placed underwater, and the dry end is placed on the shore to form a fixed small target detection sonar system. A linear frequency modulation (LFM) signal with a center frequency of 100KHz, a bandwidth of 10KHz, a pulse width of 10ms, and a repetition period of 1s is used as the detection signal. According to the different types of small targets to be detected, frogman targets and UUV targets are selected to gradually approach the detection sonar from 600m away, and surface boats are selected as other targets to gradually approach the detection sonar from 600m away. The schematic diagram of small target detection is shown in Figure 5.

[0114] Tracking trajectory extraction

[0115] In this measured data, both the training and test sets use the echo signals collected by the small target detection sonar and use the classic multiple hypothesis tracking method to obtain trajectory data. Figure 6 shows the imaging and tracking results of the small target. Among them, Figure 6 (a) is a single-frame detection output, and the bright spot in the white circle represents the moving target; Figure 6 (b) is the tracking trajectory obtained using the Multiple Hypothesis Tracking (MHT) method.

[0116] In the data, underwater slow small targets include frogmen and UUVs, and other targets are mainly surface ships and some unidentified interference. Therefore, the measured tracking tracks are all composed of frogman tracking tracks, UUV tracking tracks, surface ship tracking tracks, and other unidentified tracks generated during the tracking process. The surface ship tracking tracks and other unidentified tracks generated during the tracking process are collectively referred to as other tracking tracks.

[0117] During the data collection process, the slow underwater small target is roughly in a state of moving from far to near. The time for each underwater movement of the target is about 700 detection cycles (each detection cycle is 1 second long), forming a long tracking trajectory. After obtaining the long tracking trajectory of the target, the long trajectory is cut into several short trajectories with lengths ranging from 5 to 15 detection cycles. Based on this, 172 frogman tracking trajectories, 156 UUV tracking trajectories, and 386 other tracking trajectories are obtained. Each short trajectory is used as each data for training and testing.

[0118] In the training set, there are only 86 frogman tracking trajectories and 78 UUV tracking trajectories. In the test set, there are 86 frogman tracking trajectories, 78 UUV tracking trajectories, and 386 other trajectories. It can be seen that the overall number of trajectories is small, forming a typical small sample classification and recognition problem. At the same time, in the three types of measured data of frogman, UUV, and other trajectories, the number of other trajectories is far greater than the number of frogman and UUV tracking trajectories, resulting in a class imbalance problem.

[0119] Performance testing and analysis

[0120] SVM, SVDD, and SVDD-SVM are used to classify the tracking trajectory of underwater slow small targets. The results are shown in Figure 7. According to the confusion matrix results in Figure 7, we can see that:

[0121] (1) Here, the SVM is a three-class SVM for frogmen, UUVs, and other targets. In this implementation example, the three-class SVM is referred to as SVM. The training process and testing process of the three-class SVM used here are:

[0122] The corresponding confusion matrix is ​​shown in Figure 7(a). It can be seen that although SVM can classify frogmen, UUVs and other targets at this time, since other class trajectories contain not only surface ship trajectories but also unknown class target trajectories, other class trajectories are confused with frogmen trajectories when other class trajectories are directly used for training. For frogmen targets, the SVM's accuracy is only 20.4%; for UUV targets, the SVM's recall rate is only 26.9%. This shows that traditional small sample multi-classification methods represented by SVM are difficult to solve the class imbalance problem.

[0123] (2) Here, SVDD is to design two different SVDDs for frogmen and UUVs respectively, and use these two SVDDs together to obtain the classification ability of frogmen and UUVs. In this implementation example, the classifier composed of these two SVDDs is referred to as SVDD for short. The training process of the SVDD used here is the same as the training process of the two SVDDs in SVDD-SVM.

[0124] SVDD effectively suppressed the interference of other target trajectories on the frogman and UUV target trajectories, and the corresponding confusion matrix is ​​shown in Figure 7(b). According to the confusion matrix in Figure 7(b), the recall rate of other categories of targets reached 97.4% and the precision rate reached 95.9%, indicating that other categories of targets were almost completely classified. However, the recall rate of frogmen was only 14.0% and the precision rate was only 15.2%; the recall rate of UUV was only 15.4% and the precision rate was only 15.2%. This shows that due to the close trajectory characteristics of frogmen and UUVs, the classification method using only SVDD has serious confusion when classifying frogmen and UUVs.

[0125] (3) The confusion matrix of the classification method using SVDD-SVM is shown in Figure 7(c). The classification performance of the joint classifier SVDD-SVM is significantly improved, with the recall rate of frogmen reaching 81.4% and the precision rate reaching 80.5%; the recall rate of UUV targets reaching 82.1% and the precision rate reaching 81.0%. The recall rate and precision rate of other types of targets are 95.9% and 96.4%, respectively. It can be seen that the classification method using trajectory features and SVDD-SVM can effectively complete the classification of frogmen, UUVs, and other trajectories generated by tracking.

[0126] In order to verify the stability of the classification method, a Monte Carlo experiment was conducted by randomly dividing the training set and the test set. The trajectory data of frogmen, UUVs, and other classes were randomly sampled to ensure that the ratio of the training set to the test set was 1:1, and then feature extraction and classification were performed on the training set and the test set. The above process was repeated 100 times independently, and the results are shown in Figure 8.

[0127] According to Figure 8(a) and Figure 8(b), for the frogman target, the recall and precision of SVDD-SVM are both above 80%. In some repeated tests, the recall and precision of SVDD-SVM even exceeded 90%. Although the recall of SVDD-SVM in Figure 8(a) is lower than that of SVM, as shown in Figure 8(b), due to the serious confusion of SVM in classifying other targets and frogmen, the precision of SVM for frogman targets is much lower than that of SVDD-SVM and has obvious numerical fluctuations.

[0128] According to Figure 8(c) and Figure 8(d), for UUV targets, the recall and precision of SVDD-SVM are mostly above 80%. In some repeated tests, the recall and precision of SVDD-SVM even exceeded 90%. Similar to the situation in Figure 8(b), Figure 8(d) reflects that SVM cannot classify other targets, resulting in the precision of UUV being much lower than that of SVDD-SVM.

[0129] The average results of 100 Monte Carlo experiments are shown in Table 3. The average recall rate of SVDD-SVM for frogman targets can reach 86%, and the average precision rate can reach 87%. The average recall rate of SVDD-SVM for UUV targets can reach 85%, and the average precision rate can reach 86%. This shows that the proposed classification method not only has high classification performance for frogmen and UUVs, but also has high stability.

[0130] According to the implementation examples, it can be seen that the underwater slow small target classification method using tracking trajectory features and joint classifiers has higher precision, recall rate and accuracy, and can alleviate the degradation of classification performance of underwater slow small targets due to unstable features, small sample data and class imbalance.

[0131] Table 1 Trajectory characteristics of underwater slow small target tracking

[0132]

[0133] Table 2 Voting criteria for SVDD-SVM joint classifier

[0134]

[0135]

[0136] Table 3 Average results of 100 Monte Carlo experiments of SVM, SVDD, and SVDD-SVM

[0137]

Claims

1. A method for classifying underwater slow small targets using trajectory features and a joint classifier, characterized in that: The following steps are involved: Step 1: Extract the tracking trajectory feature of underwater slow small target; Step 2: Design SVDD-SVM joint classifier: This joint classifier adopts a two-level classifier series architecture; the first-level classifier contains two parallel SVDDs, and the second-level classifier contains a trained SVM, which is connected in series after the two SVDDs; Step 3: Train and test the SVDD-SVM joint classifier; Step 4: Design the voting criteria required for classification for the trained and tested SVDD-SVM joint classifier: If the two SVDDs obtain mutually supportive results, then directly use the result as the final classification result; If the two SVDDs obtain contradictory results, the result of the SVM classifier is taken as the final classification result; Step 5: For the tracking trajectory to be classified, extract the feature quantity according to step 1, and use the SVDD-SVM joint classifier trained and tested in step 3 to classify the tracking trajectory according to the voting criteria in step 4, so as to achieve the classification of underwater slow small targets.

2. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 1 is characterized in that: In step 1, for the local feature quantity of the underwater slow small target, the minimum value, maximum value, mean value and variance are extracted as the feature quantity; for the global feature quantity of the underwater slow small target, it can be directly used as the feature quantity, and the sum of the two types of feature quantities is used as the underwater slow small target tracking trajectory feature quantity.

3. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 1, characterized in that: In step 3, the tracking trajectory is used as training data, the feature quantity is extracted according to step 1 to construct a training set, and the SVDD-SVM joint classifier is trained, and the SVDD and SVM are trained separately.

4. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 3 is characterized in that: When training SVDD, the frogman tracking trajectory is used as training to extract feature quantities, build a training set, and train SVDD No. 1, whose output results are "frogman" and "others"; The UUV tracking trajectory used as training is used to extract feature quantities, construct a training set, and train SVDD No. 2, whose output results are "UUV" and "Others".

5. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 3, characterized in that: When training the SVM, the SVM can classify the confused data in the confused area of ​​the SVDD classifier, and the output results are "frogman" and "UUV".

6. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 1, characterized in that: In step 3, the tracking trajectory is used as the test data, and the feature quantity is extracted according to step 1 to construct a test set to test the SVDD-SVM joint classifier; during the test, the final classification output result is given by the output results of the two SVDDs and SVM and the voting criteria.

7. The underwater slow small target classification method using trajectory features and a joint classifier as claimed in claim 1, characterized in that: In step 4, the voting criteria specifically include the following steps: Step 4.1: Send the test data to the two first-level SVDDs synchronously. When the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "Others", classify the data as "Frogman"; when the output of SVDD No. 1 is "Others" and the output of SVDD No. 2 is "UUV", classify the data as "UUV"; when the outputs of SVDD No. 1 and SVDD No. 2 are both "Others", classify the data as "Others"; when the output of SVDD No. 1 is "Frogman" and the output of SVDD No. 2 is "UUV", send the contradictory test data to the second-level SVM; Step 4.2: Take the output of the second-level single SVM as the output of the joint classifier: when the SVM output is "frogman", the data is classified as "frogman"; when the SVM output is "UUV", the data is classified as "UUV".

Images