Sonar image target tracking method, target tracking system and computing device

By employing a sonar image target tracking method, which combines random cropping and deep learning model matching of feature maps from previous and subsequent frames, the false alarm problem in target tracking of sonar images is solved. This method enables effective tracking of targets with changing shapes and intensity, improving robustness and computational speed.

CN118644526BActive Publication Date: 2024-11-08BAOLI TIANTONG INSTR CORP +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411123499.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-08
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

Existing sonar image target tracking methods are prone to false alarms when faced with clutter interference and unfixed backgrounds, and are difficult to effectively track targets that move slowly or move back and forth within a small range.

Method used

A sonar image target tracking method is adopted. By randomly splitting continuous data into multiple frames, feature maps are generated and dynamic threshold segmentation is set to build a target tracking model. A deep learning model is used to match the feature maps of the previous and next frames. Background judgment is performed by combining global, local and object saliency maps. The target tracking model is then trained to improve robustness.

Benefits of technology

It improves the robustness of target tracking in sonar images, reduces false alarms, and can effectively track targets with changes in shape and intensity, while requiring less computation and being faster.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118644526B_ABST
    Figure CN118644526B_ABST
Patent Text Reader

Abstract

The invention discloses a target tracking method, a target tracking system and a computing device for a sonar image, and relates to the technical field of underwater target tracking. The target tracking method comprises: randomly dividing continuous data collected by sonar to form at least two multi-frame continuous data segments; performing visualization processing on each continuous data segment and storing it in an image format; generating a feature map according to a beam map; setting a dynamic threshold according to a pixel intensity range of the feature map, segmenting the feature map with the dynamic threshold to obtain a feature segmentation map; taking a pixel with the highest pixel intensity in a connected area as a target position; building and training a target tracking model; performing target detection on a single-frame sonar image in continuous data, inputting the segmentation feature maps of the previous frame and the current frame into a target tracking model to obtain two model output images, searching for targets with the same label, obtaining the target corresponding label in the target detection result according to the pixel position in the model output image, and realizing target tracking in continuous data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of underwater target tracking, and in particular to a sonar image target tracking method, a target tracking system and a computing device. Background Art

[0002] The existing technology mainly uses the frame difference method to subtract continuous sonar images to obtain the changed part in the stable image. The frame difference method can effectively track the target moving in the same direction stably, but if there is clutter interference in the sonar image, the background is not fixed, or the installation method is not stable enough, the frame difference method will introduce a large number of false alarms; and, if the target moves too slowly or moves in a reciprocating manner in a small range, the frame difference method is likely to lose part of the target shape and cannot effectively judge the target. Summary of the invention

[0003] The present invention provides a sonar image target tracking method, a target tracking system and a computing device, wherein the target tracking method is used to track a target in continuous data.

[0004] According to one aspect of the present invention, the present invention provides a method for tracking a target using a sonar image, wherein the method comprises the following steps:

[0005] Step 1, randomly dividing the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments;

[0006] Step 2, visualize each continuous data segment and store it in the data format of an image;

[0007] Step 3, generating a feature map according to the beam map;

[0008] Step 4, after setting a dynamic threshold according to the pixel intensity range of the feature map, segment the feature map with the dynamic threshold, and retain the part of the feature map with pixel intensity higher than the dynamic threshold as the feature segmentation map;

[0009] Step 5, taking the pixel with the highest pixel intensity in the connected area as the target position;

[0010] Step 6, building a target tracking model, the process includes: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, each convolution module is a residual structure; third, the convolution feature maps obtained by the two input branches through the convolution modules are channel-joined, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at this time, the pixel value in the model output image represents the label of the target to which the pixel position belongs, when the two model input images have a pixel position with the same label, the pixel position is regarded as the same target to complete the construction of the target tracking model;

[0011] Step 7, training the target tracking model, the process includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, annotating the continuous data, annotating the same target in the continuous data with the same digital label greater than 0, and generating an annotated image based on the annotation result, in which the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into proportions; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to the cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model;

[0012] Step 8. After loading the target tracking model and the model parameters, perform target detection on a single-frame sonar image in the continuous data, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same label on the two model output images, obtain the corresponding target label in the target detection result according to the pixel position in the model output image, and realize target tracking in the continuous data.

[0013] Preferably, the step 3 further comprises:

[0014] Step 3.1, calculate the global saliency map, local saliency map and object saliency map of the beam map;

[0015] Step 3.2, after respectively estimating the background intensities of the global saliency map, the local saliency map, and the object saliency map, background intensity maps of the global saliency map, the local saliency map, and the object saliency map are obtained;

[0016] Step 3.3, respectively calculate the logarithm of the ratio of the pixel values ​​of the global saliency map, the local saliency map, and the object saliency map to the background intensity value estimated by itself, and obtain the intensity map based on its own background;

[0017] Step 3.4: Normalize the intensity maps of the global saliency map, local saliency map, and object saliency map to , and then accumulate them to generate a feature map.

[0018] Preferably, in step 3.2, the step of estimating the background intensity value of the global saliency map is:

[0019] First, calculate the standard deviation and mean of the global saliency map;

[0020] Secondly, the global saliency map is evenly divided into four sub-maps;

[0021] Third, calculate the standard deviation for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the global saliency map, calculate the mean of the sub-image. When the mean of the sub-image is lower than 1 / 2 of the mean of the global saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the global saliency map, continue to evenly divide the sub-image into four new sub-images, and continue to calculate the standard deviation and mean of the new sub-image to determine whether the new sub-image is the background. If the new sub-image is the background, mark the new sub-image area. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the global saliency map, continue to evenly divide the new sub-image and make corresponding judgments until the size of the divided sub-image is no more than 10 pixels. At this time, directly calculate the mean of the sub-image with a size no more than 10 pixels and determine whether it is the background.

[0022] Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the global saliency map.

[0023] Preferably, in step 3.2, the step of estimating the background intensity value of the local saliency map is:

[0024] First, the standard deviation and mean of the local saliency map are calculated;

[0025] Secondly, the local saliency map is evenly divided into four sub-maps;

[0026] Third, the standard deviation is calculated for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the local saliency map, the mean of the sub-image is calculated. When the mean of the sub-image is lower than 1 / 2 of the mean of the local saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the local saliency map, the sub-image is further divided into four new sub-images, and the standard deviation and mean of the new sub-image are further calculated to determine whether the new sub-image is the background. If the new sub-image is the background, the new sub-image area is marked. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the local saliency map, the new sub-image is further divided evenly and a corresponding judgment is made until the size of the divided sub-image is no more than 10 pixels. At this time, the mean of the sub-image with a size no more than 10 pixels is directly calculated to determine whether it is the background.

[0027] Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the local saliency map.

[0028] Preferably, in step 3.2, the step of estimating the background intensity value of the object saliency map is:

[0029] First, calculate the standard deviation and mean of the object saliency map;

[0030] Secondly, the object saliency map is evenly divided into four sub-images;

[0031] Third, calculate the standard deviation for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the object saliency map, calculate the mean of the sub-image. When the mean of the sub-image is lower than 1 / 2 of the mean of the object saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the object saliency map, continue to evenly divide the sub-image into four new sub-images, and continue to calculate the standard deviation and mean of the new sub-image to determine whether the new sub-image is the background. If the new sub-image is the background, mark the new sub-image area. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the object saliency map, continue to evenly divide the new sub-image and make corresponding judgments until the size of the divided sub-image is no more than 10 pixels. At this time, directly calculate the mean of the sub-image with a size no more than 10 pixels and determine whether it is the background.

[0032] Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the object saliency map.

[0033] Preferably, in step 1, the number of frames of the divided continuous data segments is 5 to 8 frames.

[0034] Preferably, in step 4, the formula for dynamic threshold setting is , where the relation is the dynamic threshold, parameter is the coefficient of the dynamic threshold, and the parameter is the average pixel value of the feature map, and the parameter is the maximum pixel value of the feature map.

[0035] According to another aspect of the present invention, the present invention further provides a target tracking system for sonar images, comprising:

[0036] A segmentation unit, used for randomly segmenting the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments;

[0037] A visualization unit, used for visualizing each continuous data segment and storing it in an image data format;

[0038] A feature map generating unit, used to generate a feature map according to the beam map;

[0039] A feature map segmentation unit, configured to segment the feature map using the dynamic threshold after setting the dynamic threshold according to the pixel intensity range of the feature map, and retain the portion of the feature map where the pixel intensity is higher than the dynamic threshold as the feature segmentation map;

[0040] A target detection unit, used to take a pixel with the highest pixel intensity in the connected area as the target position;

[0041] The model building unit is used to build a target tracking model, and the process includes: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, and each convolution module is a residual structure; third, channel splicing is performed on the convolution feature maps obtained by the two input branches through the convolution modules, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at which time the pixel value in the model output image represents the label of the target to which the pixel position belongs. When the two model input images have a pixel position with the same label, the pixel position is regarded as the same target to complete the construction of the target tracking model;

[0042] The model training unit is used to train the target tracking model, and the process includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, the continuous data is annotated, and the same target in the continuous data is annotated with the same digital label greater than 0, and a labeled image is generated based on the annotation result, in which the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into a proportion; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to a cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model;

[0043] The target tracking unit is used to perform target detection on a single-frame sonar image in continuous data after loading the target tracking model and the model parameters, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same labels on the two model output images, obtain the corresponding target labels in the target detection results according to the pixel positions in the model output images, and realize target tracking in continuous data.

[0044] Preferably, the process of generating a feature map according to the beam map by the feature map generating unit includes: calculating a global saliency map, a local saliency map and an object saliency map of the beam map; obtaining background intensity maps of the global saliency map, the local saliency map and the object saliency map after respectively estimating the background intensities of the global saliency map, the local saliency map and the object saliency map; respectively calculating the logarithm of the ratio of the pixel values ​​of the global saliency map, the local saliency map and the object saliency map to the background intensity values ​​estimated by themselves, and obtaining an intensity map based on their own background; respectively normalizing the intensity maps of the global saliency map, the local saliency map and the object saliency map to , and then accumulate them to generate a feature map.

[0045] According to another aspect of the present invention, the present invention further provides a computing device, comprising a processor and a memory coupled to the processor, wherein the memory is used to store codes, and the processor is used to execute the codes. When the codes are executed, the processor executes a target tracking method, wherein the target tracking method comprises the following steps:

[0046] Step 1, randomly dividing the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments;

[0047] Step 2, visualize each continuous data segment and store it in the data format of an image;

[0048] Step 3, generating a feature map according to the beam map;

[0049] Step 4, after setting a dynamic threshold according to the pixel intensity range of the feature map, segment the feature map with the dynamic threshold, and retain the part of the feature map with pixel intensity higher than the dynamic threshold as the feature segmentation map;

[0050] Step 5, taking the pixel with the highest pixel intensity in the connected area as the target position;

[0051] Step 6, building a target tracking model, the process includes: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, each convolution module is a residual structure; third, the convolution feature maps obtained by the two input branches through the convolution modules are channel-joined, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at this time, the pixel value in the model output image represents the label of the target to which the pixel position belongs, when the two model input images have a pixel position with the same label, the pixel position is regarded as the same target to complete the construction of the target tracking model;

[0052] Step 7, training the target tracking model, the process includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, annotating the continuous data, annotating the same target in the continuous data with the same digital label greater than 0, and generating an annotated image based on the annotation result, in which the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into proportions; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to the cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model;

[0053] Step 8. After loading the target tracking model and the model parameters, perform target detection on a single-frame sonar image in the continuous data, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same label on the two model output images, obtain the corresponding target label in the target detection result according to the pixel position in the model output image, and realize target tracking in the continuous data.

[0054] Compared with the prior art, the present invention has at least the following beneficial effects:

[0055] 1. The target tracking method of the present invention calculates the feature map based on the sonar beam map. Compared with the general deep learning target tracking method, the target tracking method of the present invention performs better in tracking the target in the sonar image;

[0056] 2. The target to be tracked in the sonar image is detected based on a single frame of sonar image, which prevents the target to be tracked from propagating errors in continuous data, that is, when an error frame appears, it will not affect the detection of subsequent frames, so that the target tracking method of the present invention is more robust when tracking targets in continuous sonar images. In other words, each frame of sonar image is detected independently, so that when an error frame appears, it will not affect the detection of subsequent frames;

[0057] 3. The target tracking method of the present invention uses a deep learning model to match the target to be tracked in the previous and next frames. It is more robust to targets such as fishing nets whose image features such as shape and intensity change during the movement process. The target tracking model uses a fully connected layer to merge the convolution feature maps of the previous and next frames to obtain a merged feature map, and then performs subsequent processing on the merged feature map, which enables the target tracking model of the present invention to better learn the target position information of the feature map;

[0058] 4. The intensity maps of the global saliency map, local saliency map, and object saliency map are all generated based on their own background intensity. Their values ​​are more general and are not easily affected by the maximum value change when the target disappears, so there will not be a large number of false alarms when there is a target-free image;

[0059] 5. The target tracking method of the present invention mainly performs background judgment on the sub-image based on the standard deviation and the mean. Compared with the filtering-based method, the calculation amount is smaller and the minimum sub-image area is controllable. Therefore, the algorithm can sacrifice accuracy to obtain faster calculation speed as needed.

[0060] Other beneficial effects of the present invention will be further illustrated in the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1It is a flow chart of a target tracking method using a sonar image according to a preferred embodiment of the present invention.

[0062] Figure 2 It is a schematic diagram of the structure of the target tracking model according to the above preferred embodiment of the present invention.

[0063] Figure 3 It is a flow chart of a target tracking method using a sonar image according to another preferred embodiment of the present invention.

[0064] Figure 4 It is a block diagram of a sonar image target tracking system according to another preferred embodiment of the present invention.

[0065] Figure 5 It is a block diagram of a computing device according to another preferred embodiment of the present invention.

[0066] In the figure:

[0067] 400, target tracking system; 410, segmentation unit; 420, visualization unit; 430, feature map generation unit; 440, feature map segmentation unit; 450, target detection unit; 460, model building unit; 470, model training unit; 480, target tracking unit;

[0068] 500, computing device; 510, processor; 520, memory; 530, transceiver. DETAILED DESCRIPTION

[0069] Before describing in detail any embodiment of the present invention, it should be understood that the present invention is not limited in its application to the construction and arrangement details of the components described below or illustrated in the following figures. The present invention can have other embodiments and can be practiced or carried out in various ways. In addition, it should be understood that the words and terms used here are for descriptive purposes and should not be considered restrictive. The use of "including" or "having" and its variations herein is intended to cover the items and their equivalents and additional items displayed below. Unless otherwise specified or limited, the terms "install", "connect", "support" and "couple" and their variations are widely used and cover direct installation and indirect installation, connection, support and connection. In addition, "connect" and "couple" are not limited to physical or mechanical connections or connections.

[0070] Furthermore, on the first aspect, in the disclosure of the present invention, the terms "longitudinal", "lateral", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore the above terms cannot be understood as limitations on the present invention; on the second aspect, the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element may be one, while in another embodiment, the number of the element may be multiple, and the term "one" cannot be understood as a limitation on the quantity.

[0071] Figure 1 FIG. 2 is a flow chart showing a method for tracking a target using a sonar image according to a preferred embodiment of the present invention. Figure 1 As shown, the target tracking method includes the following steps:

[0072] S1, Dataset organization.

[0073] First, the sonar (for example, a forward-looking sonar) is fixed in the water and then collected. The underwater target is towed by a boat in different motion modes, and the sonar data is stored in the form of continuous data. Secondly, the complete continuous data is randomly divided into at least two continuous data segments of multiple frames. Thirdly, the sonar data of each continuous data segment is visualized and stored in the data format of an image. For example, in a specific example of the target tracking method of the present invention, the complete continuous data can be randomly divided into multiple continuous data segments of 5 to 8 frames when collating the data set. In a specific example of the target tracking method of the present invention, the sonar data of each continuous data segment can be visualized frame by frame when collating the data set, and stored in the data format of an image.

[0074] S2, feature map generation.

[0075] First, the global saliency map, local saliency map and object saliency map of the beam map are calculated. Secondly, the background intensity of the global saliency map, local saliency map and object saliency map is estimated respectively, and the background intensity values ​​of the global saliency map, local saliency map and object saliency map are calculated. Thirdly, the logarithm of the ratio of the pixel value of the global saliency map, local saliency map and object saliency map to the estimated background intensity value is calculated respectively to obtain the intensity map based on the background itself. The three intensity maps are normalized to Within the range of , a feature map is accumulated to generate a feature map.

[0076] In a specific example of the target tracking method of the present invention, the target tracking method calculates the contrast level of pixels in the beam map in the global world, and the pixel is differentially enhanced with the global level to be recorded as a global saliency map. The target tracking method calculates the local contrast level of pixels in the original image data of the sonar image, and the local mean is differentially enhanced with the global level to be recorded as a local saliency map. The target tracking method sets an object detection convolution kernel group, wherein the object detection convolution kernel group includes an object buffer part and a background estimation part, and the object detection convolution kernel group is used to perform a convolution operation on the image data of the beam map, and the maximum value of the component is taken at each pixel position in the obtained convolution image group to obtain the object saliency map. It can be seen that in some embodiments of the target tracking method of the present invention, the intensity maps of the global saliency map, the local saliency map, and the object saliency map are all generated based on their own background intensity, and their values ​​are more general and not easily affected by the change of the maximum value when the target disappears, so that a large number of false alarms will not occur when there is a target-free image.

[0077] In a specific example of the target tracking method of the present invention, first, the standard deviation and mean of the global saliency map are calculated; second, the global saliency map is evenly divided into four sub-images; third, the standard deviation is calculated for each sub-image, if the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the global saliency map, the mean of the sub-image is calculated, when the mean of the sub-image is lower than 1 / 2 of the mean of the global saliency map, the sub-image is considered to be the background, and the sub-image area is marked, if the standard deviation of the sub-image is not lower than 1 / 3 of the global saliency map, the sub-image is further evenly divided into four sub-images. Divide into four new sub-images, and continue to calculate the standard deviation and mean of the new sub-image to determine whether the new sub-image is the background. If the new sub-image is the background, mark the new sub-image area. If the standard deviation of the new sub-image is still not less than 1 / 3 of the global saliency map, continue to divide the new sub-image evenly and make corresponding judgments until the size of the divided sub-image is no more than 10 pixels. At this time, directly calculate the mean of the sub-image with a size of no more than 10 pixels and determine whether it is the background; Fourth, calculate the pixel mean for all parts marked as background, and record it as the background intensity value of the global saliency map. Correspondingly, through the above steps, the target tracking method of the present invention can obtain the background intensity value of the local saliency map and the background intensity value of the object saliency map respectively. It can be seen that the target tracking method of the present invention mainly judges the background of the sub-image based on the standard deviation and mean. Compared with the filtering-based method, the calculation amount is smaller, and the minimum sub-image area is controllable. Therefore, the algorithm can sacrifice accuracy to obtain a faster calculation speed as needed.

[0078] S3, object detection.

[0079] A dynamic threshold is set according to the pixel intensity range of the feature map and the feature map is segmented using the dynamic threshold. The part of the feature map with pixel intensities below the dynamic threshold is discarded, and the part above the dynamic threshold is retained. A connected region analysis is performed on the segmentation results, and the pixel with the highest pixel intensity in the connected region is taken as the target position.

[0080] In a specific example of the present invention, the formula for setting the dynamic threshold is , where the relation is the dynamic threshold, parameter is the coefficient of the dynamic threshold, parameter is the average pixel value of the feature map, and the parameter is the maximum pixel value of the feature map. For example, in a specific example of the target tracking method of the present invention, first, a set of continuous beam maps obtained by forward-looking sonar detection is input, and the beam map contains small moving targets such as iron balls towed by ships and / or swimming frogmen, and the beam map is uploaded and recorded at 5Hz. Secondly, a frame of the beam map is processed to calculate the feature map of the frame beam map. Thirdly, the dynamic threshold coefficient is set to 0.3, the feature map is segmented, and then a connected area analysis is performed in the segmentation result. After discarding non-target areas that are too large, the feature segmentation map at this time is retained, and the pixel maximum position of the corresponding beam map in each remaining connected area is recorded respectively, which is recorded as the target position.

[0081] S4, target tracking model construction and training.

[0082] Reference Figure 2 , the steps for building the target tracking model specifically include: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, wherein each convolution module is a residual structure, and each includes two 3×3 continuous convolution layers, a regularization layer and a Relu activation function; third, channel splicing is performed on the convolution feature maps obtained by the two input branches through the convolution modules, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; finally, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at which time the pixel value in the model output image represents the label of the target to which the pixel position belongs. When there is a pixel position with the same label in the two model input images, the pixel position is regarded as the same target, and the construction of the target tracking model is completed at this time.

[0083] The training steps of the target tracking model specifically include: first, based on the segmentation result of step S3, the continuous data is labeled, and the same target in the continuous data is labeled with the same digital label greater than 0, and a labeled image is generated based on the labeling result. In the labeled image, the pixel value of the background part is 0, and the pixel values ​​of the target part obtained by segmentation are all the digital labels of the target; secondly, the continuous data segments obtained by segmentation are processed in shuffle order, and the training set and the test set are divided into two sets in proportion; thirdly, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, and the output of the target tracking model is compared with the labeled image, and the loss value of the target tracking model is calculated, wherein the loss function is set to the cross entropy loss function; fourthly, the target tracking model is optimized by the stochastic gradient descent method, and the number of training times can be but not limited to 1000 times, and the model parameters when the loss value of the loss function is the lowest are retained, and the training of the target tracking model is completed at this time. In a specific example of the target tracking method of the present invention, after the continuous data segments obtained by cutting are shuffled, the training set and the test set can be divided into a ratio of 7:3. However, in other examples of the target tracking method of the present invention, the training set and the test set can also be divided into a ratio of 8:2, 9:1, etc.

[0084] S5, the target tracking model is used.

[0085] The target tracking model is loaded, and the model parameters trained in step S4 are loaded, target detection is performed on a single-frame sonar image in continuous data, the segmentation feature map of the previous frame and the segmentation feature map of the current frame are input into the target tracking model to obtain two model output images, and targets with the same label are searched on the two model output images, and the corresponding target label is obtained in the target detection result according to the pixel position in the model output image, so as to realize target tracking in continuous data.

[0086] Reference Figure 3 According to another aspect of the present invention, the present invention also provides a target tracking method of a sonar image, wherein the target tracking method comprises the following steps:

[0087] Step 310, randomly dividing the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments;

[0088] Step 320, visualizing each continuous data segment and storing it in an image data format;

[0089] Step 330, generating a feature map according to the beam map;

[0090] Step 340, after setting a dynamic threshold according to the pixel intensity range of the feature map, segment the feature map using the dynamic threshold, and retain the portion of the feature map where the pixel intensity is higher than the dynamic threshold as the feature segmentation map;

[0091] Step 350, taking a pixel with the highest pixel intensity in the connected area as the target position;

[0092] Step 360, building a target tracking model, wherein the process of building the target tracking model includes: first, setting two parallel input branches, one input branch inputting the feature segmentation map of the previous frame, and the other input branch inputting the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, and each convolution module is a residual structure; third, channel splicing is performed on the convolution feature maps obtained by the two input branches through the convolution modules, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at this time, the pixel value in the model output image represents the label of the target to which the pixel position belongs, and when the two model input images have a pixel position with the same label, the pixel position is regarded as the same target to complete the construction of the target tracking model;

[0093] Step 370, training the target tracking model, wherein the process of training the target tracking model includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, annotating the continuous data, annotating the same target in the continuous data with the same digital label greater than 0, and generating an annotated image based on the annotation result, in which the pixel value of the background part in the annotated image is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into a training set and a test set in proportion; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to a cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model;

[0094] Step 380, after loading the target tracking model and the model parameters, perform target detection on a single frame sonar image in the continuous data, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same label on the two model output images, obtain the corresponding target label in the target detection result according to the pixel position in the model output image, and realize target tracking in the continuous data.

[0095] The target tracking method of the present invention has at least the following beneficial effects:

[0096] First, the target tracking method of the present invention calculates a feature map based on the sonar beam map. Compared with general deep learning target tracking methods, the target tracking method of the present invention performs better in tracking targets in sonar images.

[0097] Secondly, the target to be tracked in the sonar image is detected based on a single frame of sonar image, which prevents the target to be tracked from propagating errors in continuous data, that is, when an error frame occurs, it will not affect the detection of subsequent frames, so that the target tracking method of the present invention is more robust when tracking targets in continuous sonar images. In other words, each frame of sonar image is detected independently, so that when an error frame occurs, it will not affect the detection of subsequent frames.

[0098] Again, the target tracking method of the present invention uses a deep learning model to match the targets to be tracked in the previous and next frames. It is more robust to targets such as fishing nets whose image features such as shape and intensity change during the movement process. The target tracking model uses a fully connected layer to merge the convolutional feature maps of the previous and next frames to obtain a merged feature map, and subsequently processes the merged feature map, which enables the target tracking model of the present invention to better learn the target position information of the feature map.

[0099] Preferably, in step 330, the step of generating a feature map according to the beam map further includes: calculating a global saliency map, a local saliency map and an object saliency map of the beam map; after respectively estimating the background intensities of the global saliency map, the local saliency map and the object saliency map, obtaining background intensity maps of the global saliency map, the local saliency map and the object saliency map; respectively calculating the logarithm of the ratio of the pixel values ​​of the global saliency map, the local saliency map and the object saliency map to the background intensity values ​​estimated by themselves, to obtain an intensity map based on their own background; respectively normalizing the intensity maps of the global saliency map, the local saliency map and the object saliency map to In this way, the intensity maps of the global saliency map, the local saliency map, and the object saliency map are all generated based on their own background intensity. Their values ​​are more general and are not easily affected by the change of the maximum value when the target disappears, so that a large number of false alarms will not occur when there is a target-free image.

[0100] Reference Figure 4According to another aspect of the present invention, the present invention also provides a target tracking system 400 for sonar images, wherein the target tracking system 400 includes a segmentation unit 410, a visualization unit 420, a feature map generation unit 430, a feature map segmentation unit 440, a target detection unit 450, a model building unit 460, a model training unit 470 and a target tracking unit 480.

[0101] The segmentation unit 410 is used to randomly segment the complete continuous data collected by the sonar to form at least two continuous data segments of multiple frames. Specifically, the sonar (for example, the forward-looking sonar) is fixed in the water for acquisition, and the underwater target is towed by a boat in different motion modes, or a frogman is allowed to swim underwater. At this time, the acquired sonar data is stored as continuous data, and then the complete continuous data is randomly segmented to form multiple continuous data segments of 5 to 8 frames.

[0102] The visualization unit 420 is used to perform visualization processing on each continuous data segment and store it in an image data format. The specific manner in which the visualization unit 420 performs visualization processing on the continuous data segments may be, but is not limited to: first, obtaining the beam pattern of a single frame of sonar data and the corresponding sampling parameters; second, performing operations such as contrast enhancement, background clutter suppression, and brightness curve control on the beam pattern to enhance the beam pattern and obtain an enhanced image of the beam pattern; third, normalizing the enhanced image to an 8-bit numerical space to obtain an 8-bit single-channel enhanced image of the beam pattern; fourth, mapping the 8-bit single-channel enhanced image of the beam pattern to 8-bit three-channel image data through a pseudo-color mapping map; fifth, calculating the pixel position of the fan diagram corresponding to the pixel position of the beam pattern according to the sampling parameters, storing the memory offset position of the enhanced image of the 8-bit three-channel beam pattern with the fan diagram coordinates as an index to generate a sonar fan conversion mapping map; sixth, reading data from the enhanced image of the 8-bit three-channel beam pattern through the memory offset in the sonar fan conversion mapping map, and copying it to the corresponding position of the fan diagram to obtain a sonar fan diagram, thereby realizing visualization processing of the continuous data segments.

[0103] The feature map generation unit 430 generates a feature image according to the beam map. In a specific example of the target tracking system 400 of the present invention, the specific process of the feature map generation unit 430 generating a feature image according to the beam map is: first, the global saliency map, local saliency map and object saliency map of the beam map are calculated, and secondly, after respectively estimating the background intensity of the global saliency map, the local saliency map and the object saliency map, the background intensity map of the global saliency map, the local saliency map and the object saliency map is obtained, and thirdly, the logarithm of the ratio of the pixel value of the global saliency map, the local saliency map and the object saliency map to the background intensity value estimated by itself is calculated, and the intensity map based on its own background is obtained, and fourthly, the intensity maps of the global saliency map, the local saliency map and the object saliency map are respectively normalized to , and then accumulate them to generate a feature map.

[0104] For example, in a specific example of the target tracking system 400 of the present invention, the feature map generation unit 430 calculates the global contrast level of pixels in the beam map, and the pixel is differentially enhanced with the global level to be recorded as a global saliency map. The feature map generation unit 430 calculates the local contrast level of pixels in the original image data of the sonar image, and the local mean is differentially enhanced with the global level to be recorded as a local saliency map. The feature map generation unit 430 sets an object detection convolution kernel group, wherein the object detection convolution kernel group includes an object buffer part and a background estimation part, and uses the object detection convolution kernel group to perform a convolution operation on the image data of the beam map, and takes the maximum value of the component at each pixel position in the obtained convolution image group to obtain an object saliency map.

[0105] The specific process of the feature map generation unit 430 estimating the background intensity value of the global saliency map may be: first, calculating the standard deviation and mean of the global saliency map; second, evenly dividing the global saliency map into four sub-maps; third, calculating the standard deviation of each sub-map respectively. If the standard deviation of the sub-map is lower than 1 / 3 of the standard deviation of the global saliency map, then calculating the mean of the sub-map. When the mean of the sub-map is lower than 1 / 2 of the mean of the global saliency map, the sub-map is considered to be the background, and the sub-map area is marked. If the standard deviation of the sub-map is not lower than 1 / 3 of the global saliency map, then continue to mark the sub-map. The sub-image is evenly divided into four new sub-images, and the standard deviation and mean of the new sub-image are continued to be calculated to determine whether the new sub-image is the background. If the new sub-image is the background, the new sub-image area is marked. If the standard deviation of the new sub-image is still not less than 1 / 3 of the global saliency map, the new sub-image is continued to be evenly divided and a corresponding judgment is made until the size of the divided sub-image is no more than 10 pixels. At this time, the mean of the sub-image with a size no more than 10 pixels is directly calculated and it is determined whether it is the background; Fourth, the pixel mean is calculated for all parts marked as background, which is recorded as the background intensity value of the global saliency map. Correspondingly, based on the same steps, the feature map generation unit 430 can also estimate the background intensity value of the local saliency map and the background intensity value of the object saliency map.

[0106] The feature map segmentation unit 440 is used to set a dynamic threshold according to the pixel intensity range of the feature map, segment the feature map with the dynamic threshold, and retain the part of the feature map with pixel intensity higher than the dynamic threshold as the feature segmentation map. The target detection unit 450 is used to take a pixel with the highest pixel intensity in the connected area as the target position.

[0107] For example, in a specific example of the target tracking system 400 of the present invention, the feature map segmentation unit 440 sets a dynamic threshold according to the pixel intensity range of the feature map and segments the feature map with the dynamic threshold. The part of the feature map with pixel intensity lower than the dynamic threshold is discarded, and the part higher than the dynamic threshold is retained. Subsequently, the target detection unit 450 performs a connected area analysis on the segmentation result and takes the pixel with the highest pixel intensity in the connected area as the target position.

[0108] In a specific example of the target tracking system 400 of the present invention, the formula for setting the dynamic threshold is: , where the relation is the dynamic threshold, parameter is the coefficient of the dynamic threshold, parameter is the average pixel value of the feature map, and the parameter is the maximum pixel value of the feature map. For example, first, a set of continuous beam maps detected by forward-looking sonar is input, and the beam map contains small moving targets such as iron balls towed by ships and / or swimming frogmen, and the beam map is uploaded and recorded at 5Hz. Secondly, a frame of the beam map is processed to calculate the feature map of the frame beam map. Thirdly, the dynamic threshold coefficient is set to 0.3, the feature map is segmented, and then the connected area analysis is performed in the segmentation result. After discarding the non-target area that is too large, the feature segmentation map at this time is retained, and the pixel maximum position of the corresponding beam map in each remaining connected area is recorded respectively, which is recorded as the target position.

[0109] The model building unit 460 is used to build a target tracking model. The specific process of the model building unit 460 building a target tracking model can be: first, set two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame. Secondly, the two input branches pass through two convolution modules respectively, and each convolution module is a residual structure. Third, the convolution feature maps obtained by the two input branches through the convolution modules are channel-joined, and a merged feature map is obtained through a 1×1 convolution layer. Fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer. Fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images. At this time, the pixel value in the model output image represents the label of the target to which the pixel position belongs. When there is a pixel position with the same label in the two model input images, the pixel position is regarded as the same target to complete the construction of the target tracking model.

[0110] The model training unit 470 is used to train the target tracking model. The specific process of the model training unit 470 training the target tracking model can be: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, the continuous data is annotated, and the same target in the continuous data is annotated with the same digital label greater than 0, and a labeled image is generated based on the annotation result. In the annotated image, the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target. Secondly, the continuous data segments obtained by cutting are processed in a disordered order, and the training set and the test set are divided into proportions. Thirdly, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, and the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to the cross entropy loss function. Fourthly, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model.

[0111] The target tracking unit 480 is used to perform target detection on a single-frame sonar image in continuous data after loading the target tracking model and loading the model parameters, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same labels on the two model output images, obtain the corresponding target label in the target detection result according to the pixel position in the model output image, and realize target tracking in continuous data.

[0112] Reference Figure 5 According to another aspect of the present invention, the present invention further provides a computing device 500, wherein the computing device 500 includes a processor 510 and a memory 520 coupled to the processor 510, the memory 520 stores codes, and the processor 510 is configured to execute the codes stored in the memory 520. In a possible implementation manner of the computing device 500 of the present invention, when the processor 510 executes the codes stored in the memory 520, the computing device 500 executes the target tracking method to achieve target tracking in continuous data.

[0113] Continue to refer to the attached Figure 5 The computing device 500 further includes a transceiver 530, wherein the transceiver 530 and the processor 510 communicate with each other through an internal connection path, and the processor 510 controls the transceiver 530 to send and / or receive signals when executing the code stored in the memory 520.

[0114] It should be understood by those skilled in the art that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

Claims

1. Target tracking method of sonar images, It is characterized in that The target tracking method comprises the following steps: Step 1, randomly dividing the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments; Step 2, visualize each continuous data segment and store it in the data format of an image; Step 3, generating a feature map according to the beam map, wherein the step 3 further comprises: step 3.1, calculating the global saliency map, local saliency map and object saliency map of the beam map; step 3.2, after respectively estimating the background intensity of the global saliency map, the local saliency map and the object saliency map, obtaining the background intensity map of the global saliency map, the local saliency map and the object saliency map; step 3.3, respectively calculating the logarithm of the ratio of the pixel value of the global saliency map, the local saliency map and the object saliency map to the background intensity value estimated by itself, and obtaining the intensity map based on its own background; step 3.4, respectively normalizing the intensity map of the global saliency map, the local saliency map and the object saliency map to , and then accumulate them to generate a feature map; Step 4, after setting a dynamic threshold according to the pixel intensity range of the feature map, segment the feature map with the dynamic threshold, and retain the part of the feature map with pixel intensity higher than the dynamic threshold as the feature segmentation map; Step 5, taking the pixel with the highest pixel intensity in the connected area as the target position; Step 6, building a target tracking model, the process includes: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, each convolution module is a residual structure; third, the convolution feature maps obtained by the two input branches through the convolution modules are channel-joined, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at this time, the pixel value in the model output image represents the label of the target to which the pixel position belongs, when the two model input images have a pixel position with the same label, the pixel position is regarded as the same target to complete the construction of the target tracking model; Step 7, training the target tracking model, the process includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, annotating the continuous data, annotating the same target in the continuous data with the same digital label greater than 0, and generating an annotated image based on the annotation result, in which the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into proportions; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to the cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model; Step 8. After loading the target tracking model and the model parameters, perform target detection on a single-frame sonar image in the continuous data, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same label on the two model output images, obtain the corresponding target label in the target detection result according to the pixel position in the model output image, and realize target tracking in the continuous data.

2. The target tracking method according to claim 1, It is characterized in that In step 3.2, the estimation step of the background intensity value of the global saliency map is: First, calculate the standard deviation and mean of the global saliency map; Secondly, the global saliency map is evenly divided into four sub-maps; Third, calculate the standard deviation for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the global saliency map, calculate the mean of the sub-image. When the mean of the sub-image is lower than 1 / 2 of the mean of the global saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the global saliency map, continue to evenly divide the sub-image into four new sub-images, and continue to calculate the standard deviation and mean of the new sub-image to determine whether the new sub-image is the background. If the new sub-image is the background, mark the new sub-image area. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the global saliency map, continue to evenly divide the new sub-image and make corresponding judgments until the size of the divided sub-image is no more than 10 pixels. At this time, directly calculate the mean of the sub-image with a size no more than 10 pixels and determine whether it is the background. Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the global saliency map.

3. The target tracking method according to claim 1, It is characterized in that In step 3.2, the estimation step of the background intensity value of the local saliency map is: First, the standard deviation and mean of the local saliency map are calculated; Secondly, the local saliency map is evenly divided into four sub-maps; Third, the standard deviation is calculated for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the local saliency map, the mean of the sub-image is calculated. When the mean of the sub-image is lower than 1 / 2 of the mean of the local saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the local saliency map, the sub-image is further divided into four new sub-images, and the standard deviation and mean of the new sub-image are further calculated to determine whether the new sub-image is the background. If the new sub-image is the background, the new sub-image area is marked. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the local saliency map, the new sub-image is further divided evenly and a corresponding judgment is made until the size of the divided sub-image is no more than 10 pixels. At this time, the mean of the sub-image with a size no more than 10 pixels is directly calculated to determine whether it is the background. Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the local saliency map.

4. The target tracking method according to claim 1, It is characterized in that In step 3.2, the estimation step of the background intensity value of the object saliency map is: First, calculate the standard deviation and mean of the object saliency map; Secondly, the object saliency map is evenly divided into four sub-images; Third, calculate the standard deviation for each sub-image. If the standard deviation of the sub-image is lower than 1 / 3 of the standard deviation of the object saliency map, calculate the mean of the sub-image. When the mean of the sub-image is lower than 1 / 2 of the mean of the object saliency map, the sub-image is considered to be the background, and the sub-image area is marked. If the standard deviation of the sub-image is not lower than 1 / 3 of the object saliency map, continue to evenly divide the sub-image into four new sub-images, and continue to calculate the standard deviation and mean of the new sub-image to determine whether the new sub-image is the background. If the new sub-image is the background, mark the new sub-image area. If the standard deviation of the new sub-image is still not lower than 1 / 3 of the object saliency map, continue to evenly divide the new sub-image and make corresponding judgments until the size of the divided sub-image is no more than 10 pixels. At this time, directly calculate the mean of the sub-image with a size no more than 10 pixels and determine whether it is the background. Fourth, the pixel mean of all parts marked as background is calculated and recorded as the background intensity value of the object saliency map.

5. The target tracking method according to any one of claims 1 to 4, It is characterized in that In step 1, the number of frames of the divided continuous data segments is 5 to 8 frames.

6. The target tracking method according to any one of claims 1 to 4, It is characterized in that In step 4, the formula for dynamic threshold setting is , where the relation is the dynamic threshold, parameter is the coefficient of the dynamic threshold, and the parameter is the average pixel value of the feature map, and the parameter is the maximum pixel value of the feature map.

7. The target tracking method according to any one of claims 1 to 4, It is characterized in that Prior to step 1, after the sonar is fixed in water, the underwater target is towed by a boat in different motion modes, and the sonar data is stored in a continuous data form.

8. Target tracking system of sonar images, It is characterized in that include: A segmentation unit, used for randomly segmenting the complete continuous data collected by the sonar to form at least two multi-frame continuous data segments; A visualization unit, used for visualizing each continuous data segment and storing it in an image data format; A feature map generating unit is used to generate a feature map according to a beam map, wherein the process of generating a feature map according to the beam map by the feature map generating unit includes: calculating a global saliency map, a local saliency map and an object saliency map of the beam map; after respectively estimating the background intensities of the global saliency map, the local saliency map and the object saliency map, obtaining background intensity maps of the global saliency map, the local saliency map and the object saliency map; respectively calculating the logarithm of the ratio of the pixel values ​​of the global saliency map, the local saliency map and the object saliency map to the background intensity values ​​estimated by themselves, obtaining an intensity map based on their own background; respectively normalizing the intensity maps of the global saliency map, the local saliency map and the object saliency map to , and then accumulate them to generate a feature map; A feature map segmentation unit, configured to segment the feature map using the dynamic threshold after setting the dynamic threshold according to the pixel intensity range of the feature map, and retain the portion of the feature map where the pixel intensity is higher than the dynamic threshold as the feature segmentation map; A target detection unit, used to take a pixel with the highest pixel intensity in the connected area as the target position; The model building unit is used to build a target tracking model, and the process includes: first, setting two parallel input branches, one input branch inputs the feature segmentation map of the previous frame, and the other input branch inputs the feature segmentation map of the next frame; second, the two input branches pass through two convolution modules respectively, and each convolution module is a residual structure; third, channel splicing is performed on the convolution feature maps obtained by the two input branches through the convolution modules, and a merged feature map is obtained through a 1×1 convolution layer; fourth, the merged feature map is one-dimensionally expanded, and the one-dimensionally expanded merged feature map is transformed into feature data of 32768 length through a linear layer; fifth, the feature data of length 32768 is two-dimensionalized into two 128×128 model output images, at which time the pixel value in the model output image represents the label of the target to which the pixel position belongs. When there is a pixel position with the same label in the two model input images, the pixel position is regarded as the same target to complete the construction of the target tracking model; The model training unit is used to train the target tracking model, and the process includes: first, based on the segmentation result obtained by segmenting the feature map with a dynamic threshold, the continuous data is annotated, and the same target in the continuous data is annotated with the same digital label greater than 0, and a labeled image is generated based on the annotation result, in which the pixel value of the background part is 0, and the pixel value of the target part obtained by segmentation is the digital label of the target; second, the continuous data segments obtained by segmentation are processed in a disorderly order, and the training set and the test set are divided into a proportion; third, the feature segmentation map corresponding to the training set is input into the target tracking model according to the previous and next frames, the output of the target tracking model is compared with the annotated image, and the loss value of the target tracking model is calculated, wherein the loss function is set to a cross entropy loss function; fourth, the target tracking model is optimized by the stochastic gradient descent method, and the model parameters when the loss value of the loss function is the lowest are retained to complete the training of the target tracking model; The target tracking unit is used to perform target detection on a single-frame sonar image in continuous data after loading the target tracking model and the model parameters, input the segmentation feature map of the previous frame and the segmentation feature map of the current frame into the target tracking model to obtain two model output images, search for targets with the same labels on the two model output images, obtain the corresponding target labels in the target detection results according to the pixel positions in the model output images, and realize target tracking in continuous data.

9. The target tracking system according to claim 8, It is characterized in that The sonar data is collected by fixing the sonar in water, using a boat to tow the underwater target in different motion patterns, and storing the sonar data in the form of continuous data.

10. computing device, It is characterized in that The computing device comprises a processor and a memory coupled to the processor, wherein the memory is used to store codes, and the processor is used to execute the codes. When the codes are executed, the computing device executes the target tracking method according to any one of claims 1 to 7.