A fractal feature-assisted method for tracking small targets in active sonar of ports

Abstract

The present invention relates to a fractal feature-assisted method for tracking weak targets in active port sonar, specifically in a clutter environment. This method, based on fractal feature assistance, establishes a probability density distribution model for the target and clutter in a fractal feature space, reconstructs the model for evaluating the confidence of trajectory assumptions in a multi-hypothesis tracking method, and establishes a fractal feature-assisted data association quality evaluation mechanism based on the original trajectory hypothesis motion score. Calculating trajectory assumption confidence within this evaluation system allows for rapid confirmation of target trajectory and robust tracking, while effectively suppressing large amounts of clutter measurements and prematurely terminating false tracks. This significantly improves computational efficiency and provides a highly efficient solution for tracking weak point targets in high-intensity, dense clutter environments.

Description

Technical Field

[0001] The present invention belongs to the field of target detection and tracking technology, and relates to a fractal feature-assisted small target tracking method for active port sonar, and more particularly to a fractal feature-assisted small target tracking method in a clutter environment. Background Art

[0002] Target tracking is the process of estimating the motion of targets in successive image frames and establishing unified labels. It is widely used in military, defense, and numerous engineering control fields. In particular, in long-range early warning detection systems such as radar reconnaissance and surface and underwater surveillance, target tracking technology is required to accurately locate and efficiently track suspected targets at sufficiently long ranges. Multi-target tracking technology is suitable for tracking maneuvering targets in complex scenarios, with multi-hypothesis tracking methods being the preferred approach.

[0003] Because the actual detection distance is far greater than the target's true size, the target echo loses most of its energy during the return process, appearing as a faint point target on the image. Furthermore, the presence of non-stationary scatterers in the detection scene can cause severe clutter interference in the image background, creating large-scale, intense background jitter that obscures the actual target measurement point. These issues can prevent the tracker from continuously tracking the target and result in a large number of erroneous track associations, which not only affects the monitoring system's ability to maintain target uniqueness but also reduces the tracker's computational efficiency, making it difficult to meet actual usage requirements. Summary of the Invention

[0004] Technical problems to be solved

[0005] In order to avoid the shortcomings of the existing technology, the present invention proposes a fractal feature-assisted port active sonar weak target tracking method, providing an efficient fractal feature-assisted weak target tracking method, which can judge the erroneous track associations that have been formed in the early stage of data association and perform early pruning processing on them, thereby achieving fast and accurate target tracking under high-intensity and dense clutter background interference.

[0006] Technical Solution

[0007] A fractal feature-assisted port active sonar weak target tracking method is characterized by the following steps:

[0008] Step 1: Perform connected area threshold detection on the echo image data collected at the current moment to obtain the position observation information corresponding to each potential target

[0009] Step 2: Based on the measured location information Calculate the probability of correlation between the current measurement and a formed trajectory as the motion score of the trajectory hypothesis:

[0010] S m (k) = S m (k-1)+ΔS m (k)

[0011]

[0012] in, represents the lth trajectory hypothesis at the k-1th (previous) moment, p(·) represents the probability density function, P D is the detection probability, λ φ is the spatial density of clutter, λ υ is the spatial density of the new target;

[0013] Step 3: For each measurement point Extract fractal features, calculate fractal eigenvalues, and add a fractal feature state vector f for each measurement point. i k ;

[0014] Step 4: Based on the Bayesian estimation theory, a track association evaluation mechanism based on fractal features is established. A fractal feature score is calculated for each track hypothesis to reflect the possibility of the existence of each potential data association.

[0015] Fractal feature score:

[0016] Among them, H0 and H1 represent the clutter hypothesis and target hypothesis respectively; Represents the use of the characteristic value f i k The measurement and trajectory of the previous moment The probability that the updated association hypothesis belongs to the target trajectory, and The probability that the updated trajectory association hypothesis belongs to clutter is to calculate the posterior probability of the clutter trajectory;

[0017] Step 5: Take the weighted sum of the motion score and fractal feature score of the trajectory hypothesis to obtain the total score of the trajectory association hypothesis:

[0018] S(k)=ω m S m (k)+ω f S f (k)

[0019] where ω m +ω f =1,ω m With ω f Sports score Sm (k) and fractal feature score S f (k) weight coefficient;

[0020] Step 6: Repeat steps 2 to 5 to calculate the motion score, feature score, and total score of each possible trajectory hypothesis at the current moment;

[0021] Step 7: Calculate the optimal global hypothesis set from all the formed trajectory hypotheses and obtain the output of the final trajectory. The process is:

[0022] Create an undirected graph G = (V, E), where V represents a vertex set, each representing a trajectory hypothesis; E represents an edge set, and an edge is generated when two trajectory hypotheses share the same measurement. Then, the total trajectory score of each trajectory hypothesis is used as the weight coefficient S of the corresponding trajectory hypothesis vertex. Solving the optimal global hypothesis is to obtain a set of independent vertices with the maximum total weight and no common edges, which can be expressed as the calculation:

[0023]

[0024] stx i +x j ≤1,(i,j)∈E

[0025] x i ∈{0,1+

[0026] Among them, x i is a decision variable, x i =1 means that the trajectory hypothesis is selected as one of the elements of the best global hypothesis, and x i =0 means excluding the hypothesis from the independent set;

[0027] Finally, x i = 1, all trajectories are assumed to be the most confirmed trajectories and are output as the target tracking results;

[0028] Step 8: Prune the trajectory hypotheses that deviate from the optimal global hypothesis.

[0029] Step 9: Repeat steps 1-9 in a new timing correlation cycle.

[0030] Step 1 obtains the position observation information corresponding to each potential target The process is as follows: According to the echo image data collected at the current time k, the connected area of the target is detected, and a cluster structure is formed for the continuous area composed of several adjacent pixels, which is used as an object representation; then the threshold detection is performed on each connected area to screen out the target object of interest, and this is used as the measurement set for target tracking Nk Represents the total number of measurements at time k; the center coordinates of each connected area are taken as the position state information of the target observation, and each measurement The corresponding position state vector is expressed as

[0031] The trajectory motion score of step 2 is as follows: based on the motion state filtering, the correlation gate range of each target trajectory is predicted, and the measurement points within the correlation gate are associated with the target trajectory; the principle for determining whether the measurement is within the correlation gate is that the statistical distance between the predicted target and the measurement is less than the preset standard distance; based on the position information of each measurement within the gate, the possibility of correlation between the current measurement and the target trajectory is calculated, and this is used as the motion score S of the trajectory. m (k).

[0032] In step 3, fractal feature extraction is performed on each measurement point using the extraction method published in the paper "Detecting moving targets in active sonar echograph of harbor environment using high-order time lacunarity".

[0033] The probability of the target trajectory The momentum update method is used to consider the cumulative contribution of the eigenvalues from the past 1→k-1 moments, while taking into account the probability that the eigenvalue measured at the current moment belongs to the target. The target trajectory probability is obtained through iterative calculation, namely:

[0034]

[0035] Among them, 0<m<1 is the momentum parameter, which is used to adjust the importance of the accumulated features and the current features; P(f i k |H1) is the feature measure f calculated under the feature statistical model of the target i k The probability of belonging to the target.

[0036] The posterior probability of the clutter trajectory is calculated It is to use the characteristic measurement f i k The corresponding clutter probability under the fractal characteristic statistical model reflects the probability that the track association belongs to the clutter hypothesis.

[0037] The fractal feature statistical model is: according to the real distribution of the target and clutter in the image in the historical echo image data, the class conditional probability density distribution of the fractal feature value in each corresponding area is obtained; then, according to the formed distribution result, the fractal feature value is obtained based on the statistical distribution model. The data is fitted with σ, γ, where σ and γ are scale parameters and shape parameters respectively. The maximum likelihood estimation method is used in the data fitting process to obtain the model parameters under the optimal estimation, thereby establishing a statistical model of the fractal characteristics of the target and clutter.

[0038] Beneficial effects

[0039] This invention proposes a fractal feature-assisted method for tracking weak targets in active port sonar environments. This method, developed in a clutter environment, uses a fractal feature-assisted approach to track weak targets. By establishing a probability density distribution model for the target and clutter in a fractal feature space, it reconstructs the model for evaluating the confidence of trajectory assumptions in multi-hypothesis tracking methods. Based on the existing trajectory hypothesis motion scores, a fractal feature-assisted data association quality evaluation mechanism is established. Calculating trajectory assumption confidence within this evaluation system allows for rapid confirmation of target trajectory and robust tracking. It also effectively suppresses large amounts of clutter measurements and prematurely terminates false tracks, significantly improving computational efficiency and providing a highly effective solution for tracking weak point targets in high-intensity, dense clutter environments.

[0040] The beneficial effects of the present invention are as follows: due to the use of a feature-assisted data association trajectory evaluation mechanism, when motion information cannot effectively distinguish between a moving target and dense clutter, additional fractal feature information can be used to increase the degree of distinction between the target and the clutter background, thereby not only being able to confirm the clutter trajectory in advance and suppress the probability of false alarms. Specifically, based on the key characteristic of fractal features that can effectively distinguish between a moving target and a high-intensity clutter background, the present invention establishes a fractal feature statistical model of the target and clutter in step 1 to determine the likelihood that a single measurement is from a target or clutter. Based on this, the present invention further accumulates the contribution of each measurement feature value in step 5 to construct a fractal feature score for a trajectory, thereby realizing an evaluation mechanism at the trajectory level that uses the fractal feature information of the data to determine whether the trajectory is from a real target. By calculating the fractal feature score of the trajectory, the real target trajectory will be highlighted with a higher fractal feature score due to its better fractal characteristics, while the false trajectory corresponding to the clutter will be suppressed due to its lower fractal feature score. This significantly overcomes the shortcomings of traditional target tracking methods that rely solely on position information for trajectory association. In particular, it addresses the difficulty of distinguishing real targets from randomly oscillating clutter based on motion scores in the presence of non-stationary clutter. Furthermore, the introduction of fractal feature scores in the trajectory score can better suppress the generation of false trajectories, significantly reducing the amount of data stored and processed by the tracker, thereby improving operational efficiency. Therefore, the present invention is suitable for application in complex, practical multi-target tracking and surveillance scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1It is a flow chart of the small target tracking method assisted by fractal features of the present invention.

[0042] Figure 2 This is a trajectory result diagram obtained by comparing the method of the present invention with the traditional target tracking method. DETAILED DESCRIPTION

[0043] The present invention will now be further described with reference to the embodiments and accompanying drawings:

[0044] Combine Figure 1 , a fractal feature-assisted dim target tracking method, the steps are as follows:

[0045] A small target tracking method based on fractal feature assistance includes the following steps:

[0046] Step 1: Based on the echo image data collected at the current time k, the connected area of the target is detected. A cluster structure is formed for the continuous area composed of several adjacent pixels, which is used as an object representation. Then, a threshold test is performed on each connected area to filter out the target object of interest, and this is used as the measurement set for target tracking. N k Represents the total number of measurements at time k. Take the center coordinates of each connected area as the position state information of the target observation, and each measurement The corresponding position state vector is expressed as

[0047] Step 2: Based on motion state filtering, predict the correlation gate range for each target trajectory and associate the measurement points within the correlation gate with the target trajectory. The principle for determining whether a measurement is within the correlation gate is that the statistical distance between the predicted target and the measurement is less than the preset standard distance. Based on the position information of each measurement within the gate, calculate the probability of correlation between the current measurement and the target trajectory, and use this as the motion score of the trajectory hypothesis. The calculation formula for the trajectory hypothesis motion score is:

[0048] S m (k) = S m (k-1)+ΔS m (k)

[0049]

[0050] in, represents the lth trajectory hypothesis at the k-1th (previous) moment, p(·) represents the probability density function, P D is the detection probability, λ φ is the spatial density of clutter, λ υ is the spatial density of the new target.

[0051] Step 3: For each measurement point Perform fractal feature extraction, calculate fractal feature values, and add a fractal feature state vector f for each measurement point i k The present invention utilizes the extraction method published in the paper "Detecting moving targets in activesonar echograph of harbor environment using high-order time lacunarity" to extract fractal features.

[0052] Step 4: Based on the Bayesian estimation theory, calculate the posterior probability ratio of whether the measurement belongs to the target or clutter, and establish a track association evaluation mechanism based on fractal characteristics, which is defined as:

[0053]

[0054] Among them, H0 and H1 represent the clutter hypothesis and target hypothesis respectively. Represents the use of the characteristic value f i k The measurement and trajectory of the previous moment The probability that the updated association hypothesis belongs to the target trajectory, and The probability that the updated trajectory association hypothesis belongs to clutter is expressed. For the former, the momentum update method is adopted, which considers the cumulative contribution of the eigenvalues at the past 1→k-1 moments, while taking into account the probability that the eigenvalue measured at the current moment belongs to the target. The target trajectory probability is obtained through iterative calculation, that is:

[0055]

[0056] Among them, 0<m<1 is the momentum parameter, which is used to adjust the importance of the accumulated features and the current features. i k |H1) is the probability that the feature measure fik belongs to the target calculated under the feature statistical model of the target.

[0057] For the calculation of the posterior probability of the clutter trajectory, based on the random distribution characteristics of the clutter, it is assumed that the clutter measurements at different times are independent of each other. It uses the characteristic measure f i k The corresponding clutter probability under the fractal feature statistical model reflects the probability that the trajectory association belongs to the clutter hypothesis. By calculating the log-likelihood ratio of the posterior probabilities of the two solved items above, we can obtain the fractal feature score of each trajectory hypothesis, thus reflecting the probability of the data association hypothesis from the perspective of fractal features.

[0058] Fractal feature statistical model: According to the real distribution of targets and clutter in the historical echo image data, the class conditional probability density distribution of the fractal feature values in the corresponding areas is obtained. Then, the statistical distribution model is used to calculate the distribution results. The data is fitted with , where σ and γ are scale and shape parameters respectively. In the data fitting process, the maximum likelihood estimation method is used to obtain the model parameters under the optimal estimation, thereby establishing a statistical model of the fractal characteristics of the target and clutter.

[0059] Step 5: Take the weighted sum of the motion score and fractal feature score of the trajectory hypothesis to get the total score of the trajectory association hypothesis. The total score of the trajectory hypothesis is calculated as:

[0060] S(k)=ω m S m (k)+ω f S f (k)

[0061] where ω m +ω f =1,ω m With ω f Sports score S m (k) and fractal feature score S f The weight coefficient of (k).

[0062] Step 6: Repeat steps 2-5 to calculate the motion score, feature score, and total score of each possible trajectory hypothesis associated with the wave gate.

[0063] Step 7: Calculate the optimal global hypothesis set from all the trajectory hypotheses that have been formed. Calculating the optimal global hypothesis set is to select a set of trajectory hypotheses that are pairwise compatible and have the highest total trajectory score from all the existing optional trajectory hypotheses. This needs to be solved by establishing a maximum weight independent set problem. First, create an undirected graph G = (V, E), where V represents a vertex set, each vertex represents a trajectory hypothesis; E represents an edge set, and when two trajectory hypotheses share the same measurement, an edge is generated. Then, the total trajectory score of each trajectory hypothesis is used as the weight coefficient S of the corresponding trajectory hypothesis vertex. Solving the optimal global hypothesis is to obtain a set of independent vertices with the largest total weight and no common edges, which can be expressed as the calculation:

[0064]

[0065] stx i +x j ≤1,(i,j)∈E

[0066] x i ∈{0,1+

[0067] Among them, x i is a decision variable, x i =1 means that the trajectory hypothesis is selected as one of the elements of the best global hypothesis, and x i =0 means excluding the hypothesis from the independent set.

[0068] Finally, x i = 1 are assumed to be the most confirmed trajectories and are output as the target tracking results.

[0069] Step 8: Prune the erroneous trajectory branches that deviate from the selected best global hypothesis. Using the standard N-scan backtracking pruning method, for each trajectory hypothesis branch, trace back to its child node at the kNth frame and truncate the trajectory hypothesis branch that deviates from the best global hypothesis at this node.

[0070] Step 9: Repeat steps 1-9 in a new timing correlation cycle.

[0071] In order to demonstrate the effect of the present invention, the effectiveness of the invention is verified by using real acquired echo image sequence data containing small moving targets. Figure 2 (a) is the tracking result obtained using the traditional multi-hypothesis tracking method. Figure 2 (bd) are the tracking results assisted by second-order, third-order, and fourth-order fractal features, respectively. It can be seen that compared with the traditional method, the multi-hypothesis tracking method assisted by fractal features can suppress a large number of false measurements while forming a complete target trajectory. Furthermore, with the gradual increase of the feature order, the probability of false alarm caused by false trajectory association is significantly reduced. In addition, the running time of the tracker is also accelerated from 1896 seconds of the traditional method to 58 seconds, 57 seconds, and 55 seconds, respectively. The method of the present invention can not only effectively maintain tracking accuracy in a dense clutter environment, but also greatly improve the computational efficiency, indicating its potential for achieving efficient and autonomous target tracking in complex environments.

Claims

1. A fractal feature-assisted port active sonar weak target tracking method, characterized in that Here are the steps: Step 1: Perform connected area threshold detection on the echo image data collected at the current moment to obtain the position observation information corresponding to each potential target Step 2: Based on the measured location information Calculate the probability of correlation between the current measurement and a formed trajectory as the motion score of the trajectory hypothesis: S m (k)=S m (k-1)+ΔS m (k) in, represents the lth trajectory hypothesis at the k-1th (previous) moment, p(·) represents the probability density function, P D is the detection probability, λ φ is the spatial density of clutter, λ υ is the spatial density of the new target; Step 3: For each measurement point Extract fractal features, calculate fractal eigenvalues, and add a fractal feature state vector f for each measurement point. i k ; Step 4: Based on the Bayesian estimation theory, a track association evaluation mechanism based on fractal features is established. A fractal feature score is calculated for each track hypothesis to reflect the possibility of the existence of each potential data association. Fractal feature score: Among them, H0 and H1 represent the clutter hypothesis and target hypothesis respectively; Represents the use of the characteristic value f i k The measurement and trajectory of the previous moment The probability that the updated association hypothesis belongs to the target trajectory, and The probability that the updated trajectory association hypothesis belongs to clutter is to calculate the posterior probability of the clutter trajectory; Step 5: Take the weighted sum of the motion score and fractal feature score of the trajectory hypothesis to obtain the total score of the trajectory association hypothesis: S(k)=ω m S m (k)+ω f S f (k) where ω m +ω f =1,ω m With ω f Sports score S m (k) and fractal feature score S f (k) weight coefficient; Step 6: Repeat steps 2 to 5 to calculate the motion score, feature score, and total score of each possible trajectory hypothesis at the current moment; Step 7: Calculate the optimal global hypothesis set from all the formed trajectory hypotheses and obtain the output of the final trajectory. The process is: Create an undirected graph G = (V, E), where V represents a vertex set, each representing a trajectory hypothesis; E represents an edge set, and an edge is generated when two trajectory hypotheses share the same measurement. Then, the total trajectory score of each trajectory hypothesis is used as the weight coefficient S of the corresponding trajectory hypothesis vertex. Solving the optimal global hypothesis is to obtain a set of independent vertices with the maximum total weight and no common edges, which can be expressed as the calculation: s.t.x i +x j ≤1,(i,j)∈E x i ∈{0,1} Among them, x i is a decision variable, x i =1 means that the trajectory hypothesis is selected as one of the elements of the best global hypothesis, and x i =0 means excluding the hypothesis from the independent set; Finally, x i = 1, all trajectories are assumed to be the most confirmed trajectories and are output as the target tracking results; Step 8: Prune the trajectory hypotheses that deviate from the optimal global hypothesis. Step 9: Repeat steps 1-9 in a new timing correlation cycle.

2. The fractal feature-assisted port active sonar small target tracking method according to claim 1 is characterized by: Step 1 obtains the position observation information corresponding to each potential target The process is as follows: According to the echo image data collected at the current time k, the connected area of the target is detected, and a cluster structure is formed for the continuous area composed of several adjacent pixels, which is used as an object representation; then the threshold detection is performed on each connected area to screen out the target object of interest, and this is used as the measurement set for target tracking N k Represents the total number of measurements at time k; the center coordinates of each connected area are taken as the position state information of the target observation, and each measurement The corresponding position state vector is expressed as 3. The fractal feature-assisted small target tracking method for active port sonar according to claim 1, characterized in that: The trajectory motion score of step 2 is as follows: based on the motion state filtering, the correlation gate range of each target trajectory is predicted, and the measurement points within the correlation gate are associated with the target trajectory; the principle for determining whether the measurement is within the correlation gate is that the statistical distance between the predicted target and the measurement is less than the preset standard distance; based on the position information of each measurement within the gate, the possibility of correlation between the current measurement and the target trajectory is calculated, and this is used as the motion score S of the trajectory. m (k).

4. The fractal feature-assisted port active sonar small target tracking method according to claim 1 is characterized by: In step 3, fractal feature extraction is performed on each measurement point using the extraction method published in the paper "Detecting moving targets inactive sonar echograph of harbor environment using high-order time lacunarity".

5. The fractal feature-assisted port active sonar small target tracking method according to claim 1 is characterized by: The probability of the target trajectory The momentum update method is used to consider the cumulative contribution of the eigenvalues from the past 1→k-1 moments, while taking into account the probability that the eigenvalue measured at the current moment belongs to the target. The target trajectory probability is obtained through iterative calculation, namely: Among them, 0<m<1 is the momentum parameter, which is used to adjust the importance of the accumulated features and the current features; P(f i k |H1) is the feature measure f calculated under the feature statistical model of the target i k The probability of belonging to the target.

6. The fractal feature-assisted port active sonar small target tracking method according to claim 1 is characterized by: The posterior probability of the clutter trajectory is calculated It is to use the characteristic measure f i k The corresponding clutter probability under the fractal feature statistical model reflects the probability that the trajectory association belongs to the clutter hypothesis.

7. The fractal feature-assisted port active sonar small target tracking method according to claim 6, characterized in that: The fractal feature statistical model is: according to the real distribution of the target and clutter in the image in the historical echo image data, the class conditional probability density distribution of the fractal feature value in each corresponding area is obtained; then, according to the formed distribution result, the fractal feature value is obtained based on the statistical distribution model. The data is fitted with σ, γ, where σ and γ are scale parameters and shape parameters respectively. The maximum likelihood estimation method is used in the data fitting process to obtain the model parameters under the optimal estimation, thereby establishing a statistical model of the fractal characteristics of the target and clutter.

Images