A tracking method for small targets with sudden size changes in forward-looking sonar images

Abstract

The present invention discloses a target tracking method that adapts to the sudden change of the size of small targets in forward-looking sonar images. It includes: setting a scale pool with a larger range; extracting target features based on the search box corresponding to the elements in the scale pool and sending them to the filter to obtain the response value; setting the threshold parameter using the ratio of the maximum value of the response value to the peak sidelobe near the maximum value; finding the initial value of the threshold parameter using the threshold parameter corresponding to all scales of the second frame image with high confidence; screening the elements in the scale pool using the initial value of the threshold parameter; constantly comparing the threshold parameter with the initial value of the threshold parameter to determine whether the target size has suddenly changed; and giving the corresponding tracking scale based on screening the elements of the scale pool. The present invention can effectively solve the problem of target tracking failure caused by the deformation of the target within a larger range, which is common in the process of forward-looking sonar target tracking, and improve the robustness of the target tracking process.

Description

[0001] Technology Neighborhood

[0002] The invention relates to the field of forward-looking sonar target tracking, and in particular to a tracking method that adapts to sudden changes in the size of small targets in forward-looking sonar images. Background Art

[0003] The characteristics of forward-looking sonar are low resolution, large changes in target features and target scale between adjacent frames. The target scale often changes during target tracking. The initial kernel correlation filter does not perform adaptive tracking based on the scale. When the target becomes smaller, the filter will learn more background information, and the accuracy of target tracking will be affected to a certain extent. More importantly, when the target scale becomes larger, the filter may only learn the local texture information of the target. At this time, the tracker is very likely to only learn to track the local texture features of the target, which will cause the failure of target tracking.

[0004] Later, three major types of target tracking methods were developed for scale adaptation of target tracking, namely, scale pool-based method, block-based method and feature point matching-based method.

[0005] Forward-looking sonar images have high image noise, so the error rate of scale adaptation using feature point matching is high; and forward-looking sonar images have many small targets, so they are not suitable for scale adaptation based on block-based methods; the scale pool-based method is more suitable for tracking small targets based on forward-looking sonar images, but the number of scale pool elements affects the running speed of the algorithm. The more the number, the slower the tracking speed, and in the process of traversing the scale pool, when the appropriate scale has been found, other elements are still traversed, resulting in a waste of computing power. Moreover, the range of the scale pool is small, which is not enough to cope with the situation where the size of the tracking target changes suddenly.

[0006] In view of the above research background, the present invention uses the maximum value of the kernel correlation filter response value to reflect the probability of the target center point at that point and the peak sidelobe ratio near the maximum value of the filter response value to reflect the degree of target deformation. Threshold parameters based on the maximum value of the filter response value and the peak sidelobe ratio near the maximum value are introduced to screen the elements in the scale pool, thereby avoiding traversing the scale pool and reducing the amount of calculation, improving the algorithm running speed, and at the same time, by estimating the sudden change of the target scale size, a sufficiently large scale is given for target search to avoid target tracking failure. Summary of the invention

[0007] The purpose of the present invention is to solve the problem of low computational efficiency of the scale pool-based method, waste of computing power caused by traversing other elements when the scale that makes tracking successful has been found; and tracking failure problem caused by sudden change of the scale of small targets in forward-looking sonar.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A tracking method adapted to sudden changes in the size of small targets in forward-looking sonar images, comprising:

[0010] Step 1: Set the scale pool and select the feature extraction range of the read current frame according to the scale corresponding to the elements in the scale pool;

[0011] Step 2: According to the order of elements in the scale pool, extract the target features one by one in different scale search boxes;

[0012] Step 3: Send the extracted target features into the kernel correlation filter to obtain the feature response value;

[0013] Step 4: Calculate the threshold parameter based on the characteristic response value, and calculate the initial value of the threshold parameter according to the threshold parameter;

[0014] Step 5: Determine whether the threshold parameter has a sudden change, and determine whether the current scale is a suitable tracking scale based on the initial value of the threshold parameter;

[0015] Step 6: Repeat steps 2 to 5 until the loop ends or all elements in the scale pool are traversed to find a suitable tracking scale;

[0016] Step 7: Get the next frame of forward-looking sonar image, repeat steps 2 to 6, calculate the adaptive tracking scale and the center point of the target location, until all forward-looking sonar images are calculated.

[0017] Preferably, step 4: calculating the threshold parameter based on the characteristic response value, and calculating the initial value of the threshold parameter according to the threshold parameter specifically includes:

[0018] Step 41: Calculate the threshold parameter St based on the characteristic response maximum value and the peak sidelobe ratio near the characteristic response maximum value. The specific calculation formula is:

[0019] St=0.3*y max +0.7*sigmoid(PSR)

[0020] Among them, 0.3 and 0.7 represent the weight values ​​corresponding to the maximum value of the characteristic response of the kernel correlation filter and the peak sidelobe ratio near the maximum value, and the peak sidelobe ratio is normalized to between 0 and 1 through the sigmoid function. PSR represents the peak sidelobe ratio, and the calculation formula is:

[0021]

[0022] y max Indicates the maximum response peak, μ sl represents the average value of the characteristic response value in the selected area, σ xIt represents the variance of the feature response value in the selected area. The area is selected as a 10*10 pixel range near the maximum value of the feature response value.

[0023] Step 42: Determine whether the current frame is the second frame. If it is the second frame and after traversing the scale pool, calculate the threshold parameter St of the feature response values ​​corresponding to all elements in the scale pool of the second frame image, and select the minimum threshold parameter as the initial value of the scale adaptive threshold parameter:

[0024] St 0 =min(St 1 ,St 2 ,St 3 ,...St n )

[0025] St 0 is the initial value of the threshold parameter, St 1 ,St 2 ,St 3 ,...St n is the threshold parameter value of the characteristic response value corresponding to each element in the scale pool, and n represents the number of elements in the scale pool;

[0026] Step 43: Select the element in the scale pool corresponding to the maximum value of the threshold parameter and multiply it by the scale of the target in the previous frame as the output tracking scale.

[0027] Preferably, step five: determining whether a threshold parameter has a sudden change, and determining whether the current scale is a suitable tracking scale according to the initial value of the threshold parameter specifically includes:

[0028] After determining that the current frame is not the second frame, starting from the first scale in the scale pool, compare the threshold parameter corresponding to the feature response value of each scale with the initial value of the threshold parameter. If it is greater than the initial value of the threshold parameter, the scale corresponding to the threshold parameter is used as the tracking scale. Otherwise, compare the threshold parameter corresponding to the element in the next scale pool with the initial value of the threshold parameter until a scale value greater than the initial value of the threshold parameter is found as the tracking scale.

[0029] Preferably, step six is ​​specifically as follows: if after traversing the entire scale pool, no threshold parameter corresponding to an element is greater than the initial value of the threshold parameter, then it is determined whether the maximum value of the threshold parameter corresponding to all elements in the scale pool is less than 80% of the initial value of the threshold parameter; if so, the target scale is selected to be increased by 1.04-1.1 times on the basis of the original scale size to cope with the tracking failure problem caused by the sudden change in the target size; otherwise, the element in the scale pool corresponding to the maximum value of the output threshold parameter is multiplied by the scale size of the target in the previous frame as the output scale.

[0030] Preferably, the specific calculation formula of the kernel correlation filter in step three is:

[0031]

[0032] Represents the Fourier transform of the result after x and z are processed by the kernel function. Represents the model parameters of the kernel correlation filter.

[0033] Preferably, the element of the scale pool is the scaling value of the target search box. For the characteristics of the small target of the forward-looking sonar, the scaling range of the scale is 1.2 times the original target scale at most and 0.8 times the original target at least. The scale interval is 0.01, and the scale of the scale pool is {1 0.99 1.01 0.98 1.02 0.97 1.03 0.96 1.04 0.95 1.05 0.94 1.06 0.931.07 0.92 1.08 1.09 0.91 1.1 0.9 1.11 0.89 1.12 0.88 1.13 0.87 1.14 0.86 1.150.85 1.16 0.84 1.17 0.83 1.18 0.82 1.19 0.81 1.2 0.8}.

[0034] Preferably, step 2 is specifically as follows: taking the target center position and target size of the current frame as a reference, taking the position of the target center point of the previous frame as the center in the next frame image, and multiplying the target size of the previous frame by 2.5 times the corresponding scale pool element as a search box to search for the target center of the next frame.

[0035] Through the above technical solutions, it can be known that compared with the existing technology, the advantages of the present invention are: the existing methods for forward-looking sonar target tracking have the problem of slow speed and low precision. For example, the forward-looking sonar target tracking method based on deep learning has the problem of slow speed; and the method based on kernel correlation filtering has a fast tracking speed, but the method using scale pool for scale adaptation has the problem of limiting the operation speed due to the number of scale pools and wasting computing power caused by traversing the scale pool; when facing the sudden change of the forward-looking sonar target, the target tracking scale corresponding to the maximum value of the filter response cannot track the target. This method does not limit the range of the scale pool, and a larger range of scale pools can be set to improve the speed while ensuring accuracy; the set threshold parameters can effectively predict the sudden change in the size of the forward-looking sonar target, set a larger scale, and search for the sudden change target in a larger range, thereby avoiding target tracking failure and improving the success rate of target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0037] Attached Figure 1 It is a flow chart of the method of the present invention.

[0038] Attached Figure 2 are the target center point and target scale box of the given initial frame.

[0039] Attached Figure 3 It is a display diagram of the search range of search boxes of different sizes.

[0040] Attached Figure 4 They are the filter characteristic response values ​​for (a) successful tracking and (b) about to fail tracking.

[0041] Attached Figure 5 It is the threshold parameter distribution map of all scales in the scale pool corresponding to the second frame.

[0042] Attached Figure 6 It is a schematic diagram of judging whether the scale is a target tracking scale based on the threshold parameter.

[0043] Attached Figure 7 This is a schematic diagram of target tracking scale selection when facing target mutation and the threshold parameters corresponding to all scales are less than 80% of the initial value of the threshold parameter. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The following will be combined Figures 2 to 7 The present invention is described in further detail.

[0046] A tracking method adapted to sudden changes in the size of small targets in forward-looking sonar images comprises the following steps:

[0047] Based on the target center point position and target size of the given initial frame, such as Figure 2 As shown, the feature extraction range is selected according to the elements in the set scale pool, and the search box position of the next frame is determined.

[0048] The elements in the scale pool are the scaling values ​​of the target search box. For the characteristics of the small target of the forward-looking sonar, the scaling range of the scale is 1.2 times the original target scale at most and 0.8 times the original target at least. The scale interval is 0.01, and the scale pool is {1 0.99 1.01 0.98 1.02 0.97 1.03 0.96 1.04 0.95 1.05 0.94 1.06 0.93 1.070.92 1.08 1.09 0.91 1.1 0.9 1.11 0.89 1.12 0.88 1.13 0.87 1.14 0.86 1.15 0.851.16 0.84 1.17 0.83 1.18 0.82 1.19 0.81 1.2 0.8}.

[0049] (2) Extract target features one by one in search boxes of different scales, such as Figure 3 Specifically, based on the target center position and target size of the current frame, the target center point of the previous frame is used as the center in the next frame image, and a search box with a size of 2.5 times the size of the corresponding scale pool element multiplied by the target size of the previous frame is used to search for the target center of the next frame.

[0050] (3) The extracted target features are sent to the kernel correlation filter to obtain the feature response value, such as Figure 4 The characteristic response value indicates the probability value of the target center point at that point. The distribution around the maximum value in the characteristic response value can be represented by the peak-to-sidelobe ratio. The change in the peak-to-sidelobe ratio can be used to determine the target tracking situation. Figure 4 (a) represents the response value of the target normal tracking, Figure 4 (b) indicates the response value when the target fails to track immediately.

[0051] (4) The threshold parameter value is calculated based on the peak sidelobe ratio of the maximum characteristic response and the peak sidelobe ratio near the maximum value. The threshold parameter is used to determine whether the current scale is a suitable target tracking scale and decide whether to end the traversal of the scale pool, thereby reducing the amount of calculation and improving the running speed, and starting to track the target of the next frame image.

[0052] The initial value of the threshold parameter is selected from the second frame, that is, the threshold parameter reference value of the target scale is selected. Because the thresholds of the initial few frames of tracking in the target tracking process have high confidence, the present invention selects the second frame. For example, the threshold parameter distribution at all scales corresponding to the second frame is as follows: Figure 5 As shown, the minimum value of the threshold parameter of the second frame is selected ( Figure 5 In the example, 0.824 is used as the initial value of the threshold parameter.

[0053] (5) Determine whether the target size has mutated, so as to estimate whether the target size has mutated.

[0054] Starting from the third frame, starting from the first element in the scale pool, the threshold parameter of the feature response value corresponding to each element is compared with the initial value of the threshold parameter. If it is greater than the initial value of the threshold parameter, it is used as the tracking scale. The schematic diagram is as follows Figure 6 As shown in the figure, the threshold parameter corresponding to the first element S1 in the scale pool is compared with the initial value of the threshold parameter. If it is greater than the initial value of the threshold parameter, the scale corresponding to the current scale pool element is output as the tracking scale; if it is less than the initial value of the threshold parameter, the threshold parameter corresponding to the next element S2 is compared with the initial value of the threshold parameter to determine whether to output the scale corresponding to the current element or to compare the threshold parameter of the next element, and so on, until a scale value greater than the initial value of the threshold parameter is found as the output.

[0055] (6) In step (5), after traversing the entire scale pool, there is no element whose corresponding threshold parameter is greater than the initial value of the threshold parameter, such as Figure 7 As shown in the figure, the maximum value of the threshold parameter corresponding to all scales is compared to see if it is less than 80% of the initial value of the threshold parameter. If it is, the scale is selected to be 1.04-1.1 times higher than the scale of the target in the previous frame to cope with the tracking failure problem caused by the sudden change of the target size; otherwise, the scale corresponding to the maximum value of the output threshold parameter is used as the output scale, that is, the output tracking scale is the element in the scale pool corresponding to the maximum value of the threshold parameter (scale scaling value) multiplied by the scale of the target in the previous frame.

[0056] (7) Repeat steps (2) to (6) to complete the tracking of the target in each frame of the forward-looking sonar.

[0057] In this embodiment, the specific method of calculating the threshold parameter of the present invention is:

[0058] First, the size of the filter response value indicates the probability that the center point of the target is at that point. The location of the maximum value of the filter response value indicates the center point of the target, and the peak sidelobe ratio near the maximum value of the filter response value can indicate the closeness between the current scale and the real target scale. That is, the closer the search box size is to 2.5 times the real target size, the larger the peak sidelobe ratio will be. The calculation formula of the peak sidelobe ratio is as follows:

[0059]

[0060] PSR stands for Peak Sidelobe Ratio, max Indicates the maximum response peak, μ sl represents the average value of the characteristic response value in the selected area, σ x Indicates the variance of the response value within the selected area. The area is selected as a 10*10 pixel range near the maximum response value.

[0061] The calculation formula of the threshold parameter St based on the maximum value of the extracted characteristic response value and the peak sidelobe ratio near the maximum value is:

[0062] St=0.3*y max +0.7*sigmoid(PSR)

[0063] 0.3 and 0.7 represent the weight values ​​corresponding to the maximum value of the filter response and the peak sidelobe ratio near the maximum value, and the peak sidelobe ratio is normalized to between 0 and 1 through the sigmoid function. max It can only reflect the confidence of the center position of the tracked target, while PSR can reflect the target tracking situation. The larger the PSR value, the closer the size of the search box is to 2.5 times the actual size of the target; the smaller the PSR value, the greater the error between the size of the search box and the actual size of the target, and the more likely it is that the target tracking will fail. max The smaller the weight, the greater the weight of PSR.

[0064] In this embodiment, the initial value of the threshold parameter is selected based on the high confidence of St corresponding to the initial frame. Combined with the first and second frames, the threshold parameter St of the response value corresponding to all scales in the scale pool of the second frame is calculated, and the smallest St is selected as the threshold for scale adaptation judgment.

[0065] St 0 =min(St 1 ,St 2 ,St 3 ,...St n )

[0066] St 0 is the initial value of the threshold parameter, is the threshold used for scale adaptive judgment, St 1 ,St 2 ,St 3 ,...St n is the threshold parameter value of the feature response value corresponding to each element in the scale pool, and n represents the number of elements in the scale pool.

[0067] In this embodiment, the kernel correlation filter is specifically implemented as follows:

[0068] Define the response value f as a linear function of the target feature x, f(x i )=W T x i . Using sample cyclic shift instead of sampling window, we can get the loss function:

[0069] loss=min w ∑(f(xi )-y i ) 2 +λ‖w‖ 2

[0070] x i is the training sample, y i is the training sample label, λ represents the regularization term, and the partial derivative of the loss function is simplified to obtain the optimal solution: w = (X T X+λI) -1 X T y. And because w can be represented by x. We get the expression for predicting the next frame of image:

[0071]

[0072] Where x i represents the training sample obtained based on the previous frame image, and z represents the next frame image. Since the kernel function technique can solve the problem that the original vector cannot linearly regress the sample label, and the kernel function can use the circulant matrix to simplify the operation, combined with the diagonal property of the circulant matrix, the expression of α can be obtained:

[0073]

[0074] λ represents the regularization term, represents Fourier transform of the label, ⊙ represents the bitwise multiplication of the matrix, It represents the Fourier transform of the kernel function output. The Gaussian kernel function is used. σ represents variance, F -1 represents the inverse Fourier transform, x and x' represent the input of the kernel function;

[0075] because Perform Fourier transform on both sides of the expression at the same time, and finally get the expression of the response value:

[0076]

[0077] Represents the Fourier transform of the result after x and z are processed by the kernel function. Represents the model parameters of the kernel correlation filter.

[0078] In this embodiment, the method for judging the sudden change of the threshold parameter and adjusting the scale selection in step (4) to avoid target tracking failure is specifically as follows:

[0079] The size of the target in the next frame of the current frame suddenly changes exponentially. The sudden change in the size of the target causes a huge change in the size of the threshold parameter. At this time, the threshold parameters corresponding to the elements in the scale pool may be smaller than the initial value of the threshold parameter. After traversing all the elements in the scale pool, calculate whether the maximum value of St corresponding to all the scale pool elements is less than St 0 80% less than St 0 80% of the target is judged as a target. Since the target size mutation is prone to tracking failure, a larger scale is specified for tracking. The scale selection range is between 1.04-1.1 (so that the search box is enlarged, so that the target can be searched in a larger range in the next frame, so that the St value is in the correct range; here a smaller scale scaling value of 1.04 is selected to enlarge the scale to avoid the search box being further enlarged and drifting when the St value returns to the correct range); otherwise, the scale corresponding to the maximum value of St is selected as the adaptive tracking scale.

[0080] When the maximum value of the threshold parameter is less than 80% of the initial value of the threshold parameter, the target scale (tracking box size) is selected to be 1.04 times higher than the original scale, and the output scale is 1.04 times the target scale of the previous frame, that is, the element 1.04 in the scale pool is selected to cope with the tracking failure problem caused by the sudden change of the target size:

[0081] S=1.04

[0082] S represents the scale scaling value. The output tracking scale is the target size of the previous frame multiplied by the scale scaling value S.

[0083] When there is a threshold parameter greater than 80% of the initial value of the threshold parameter:

[0084] S=S max(St)

[0085] S max(St) Indicates the scale scaling value in the scale pool corresponding to the maximum value of St.

[0086] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0087] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A tracking method that adapts to the sudden change of the size of small targets in forward-looking sonar images, characterized by: include: Step 1: Set the scale pool and select the feature extraction range of the read current frame according to the scale corresponding to the elements in the scale pool; Step 2: According to the order of elements in the scale pool, extract the target features one by one in different scale search boxes; Step 3: Send the extracted target features into the kernel correlation filter to obtain the feature response value; Step 4: Calculate the threshold parameter based on the characteristic response value, and calculate the initial value of the threshold parameter according to the threshold parameter; Specifically include: Step 41: Calculate the threshold parameter St based on the characteristic response maximum value and the peak sidelobe ratio near the characteristic response maximum value. The specific calculation formula is: St=0.3*y max +0.7*sigmoid(PSR) Among them, 0.3 and 0.7 represent the weight values ​​corresponding to the maximum value of the characteristic response of the kernel correlation filter and the peak sidelobe ratio near the maximum value, and the peak sidelobe ratio is normalized to between 0 and 1 through the sigmoid function. PSR represents the peak sidelobe ratio, and the calculation formula is: y max Indicates the maximum response peak, μ sl represents the average value of the characteristic response value in the selected area, σ x It represents the variance of the feature response value in the selected area. The area is selected as a 10*10 pixel range near the maximum value of the feature response value. Step 42: Determine whether the current frame is the second frame. If it is the second frame and after traversing the scale pool, calculate the threshold parameter St of the feature response values ​​corresponding to all elements in the scale pool of the second frame image, and select the minimum threshold parameter as the initial value of the scale adaptive threshold parameter: St0=min(St1,St2,St3,...St n ) St0 is the initial value of the threshold parameter, St1, St2, St3, ...St n is the threshold parameter value of the characteristic response value corresponding to each element in the scale pool, and n represents the number of elements in the scale pool; Step 43: Select the element in the scale pool corresponding to the maximum value of the threshold parameter and multiply it by the scale of the target in the previous frame as the output tracking scale; Step 5: Determine whether the threshold parameter has a sudden change, and determine whether the current scale is a suitable tracking scale based on the initial value of the threshold parameter; Step 6: Repeat steps 2 to 5 until the loop ends or all elements in the scale pool are traversed to find a suitable tracking scale; Step 7: Get the next frame of forward-looking sonar image, repeat steps 2 to 6, calculate the adaptive tracking scale and the center point of the target location, until all forward-looking sonar images are calculated.

2. The tracking method according to claim 1, which is adapted to the sudden change of the size of small targets in forward-looking sonar images, is characterized in that: Step 5: Determine whether the threshold parameter has a sudden change, and determine whether the current scale is an appropriate tracking scale based on the initial value of the threshold parameter. Specifically include: After determining that the current frame is not the second frame, starting from the first scale in the scale pool, compare the threshold parameter corresponding to the feature response value of each scale with the initial value of the threshold parameter. If it is greater than the initial value of the threshold parameter, the scale corresponding to the threshold parameter is used as the tracking scale. Otherwise, compare the threshold parameter corresponding to the element in the next scale pool with the initial value of the threshold parameter until a scale value greater than the initial value of the threshold parameter is found as the tracking scale.

3. The tracking method according to claim 2, which is adapted to the sudden change of the size of small targets in forward-looking sonar images, is characterized in that: Step 6 is as follows: If after traversing the entire scale pool, no element corresponds to a threshold parameter greater than the initial value of the threshold parameter, then determine whether the maximum value of the threshold parameters corresponding to all elements in the scale pool is less than 80% of the initial value of the threshold parameter. If it is less, the scale is selected to be 1.04-1.1 times higher than the target scale of the previous frame to cope with the tracking failure problem caused by the sudden change in target size; otherwise, the element in the scale pool corresponding to the maximum value of the output threshold parameter is multiplied by the scale of the target of the previous frame as the output scale.

4. The tracking method for small targets in forward-looking sonar images adapted to sudden changes in size according to claim 1, characterized in that: Step 3: The specific calculation formula of the kernel correlation filter is: Represents the Fourier transform of the result after x and z are processed by the kernel function. Represents the model parameters of the kernel correlation filter.

5. The tracking method for small targets in forward-looking sonar images adapted to sudden changes in size according to claim 1, characterized in that: The elements of the scale pool are the scaling values ​​of the target search box. For the characteristics of small targets in the forward-looking sonar, the scaling range of the scale is 1.2 times the original target scale at most and 0.8 times the original target at least. The scale interval is 0.01, and the scale of the scale pool is {10.99 1.01 0.98 1.02 0.97 1.03 0.96 1.04 0.95 1.05 0.94 1.06 0.93 1.07 0.921.08 1.09 0.91 1.1 0.9 1.11 0.89 1.12 0.88 1.13 0.87 1.14 0.86 1.15 0.85 1.160.84 1.17 0.83 1.18 0.82 1.19 0.81 1.2 0.8}.

6. The tracking method for small targets in forward-looking sonar images adapted to sudden changes in size according to claim 1, characterized in that: Step 2 is as follows: based on the target center position and target scale of the current frame, the target center point of the previous frame is used as the center in the next frame image, and the target size of the previous frame multiplied by 2.5 times the size of the corresponding scale pool element is used as the search box to search for the target center of the next frame.

Images