Method and device for recognizing a maritime target

Abstract

The application discloses a method and device for identifying a marine target, the method comprises the following steps: performing target detection on signal sources of multiple sensors respectively to obtain multiple detection results; recording the multiple detection results in a unified preset form; aligning coordinate systems of the multiple sensors according to calibration of the multiple sensors to obtain a new scene representation space; projecting the multiple detection results into the new scene representation space respectively to obtain corresponding multiple new detection results; fusing the multiple new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in a candidate target set; and performing target identification on all the fusion detection results according to a multi-data fusion target identification algorithm to determine a target type of each target. The application solves the problem that the existing detection method for small marine targets has a large error and influences the accuracy of small target identification in a marine scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of territorial sea safety, in particular to a sea target identification method and device. BACKGROUND

[0002] With the rapid development of navigation technology and detection technology, marine resources have increasingly become one of the main resources that countries compete for. Under the above background, sea target identification and track prediction are important links for many practical applications. Good performance of target identification and track prediction is of great significance for fishery management, marine pollution monitoring and monitoring, air and sea accident search and rescue, illegal ship expulsion, etc. In order to improve the efficiency of sea navigation and maintain the safety of the sea area, it provides a solid foundation. In summary, whether in military or civilian, the importance of sea target identification and track tracking is self-evident. However, due to the complexity of the sea application scene, the identification and track prediction technology of the sea target, especially the small sea target, is not perfect. From the complexity of the sea scene, due to the characteristics of the marine environment and the sea target and the difference between the sea target and the on-road target such as pedestrians and vehicles, the existing classic target identification, tracking and track prediction technology is difficult to directly reuse and produce good results.

[0003] From the current detection technology for small sea targets, although the small target detection problem has become a research hotspot in the application fields of fishery management, national defense safety, etc., many domestic and foreign research institutions have also carried out extensive research on it. However, due to the characteristics of the small target itself and the defects of the detection equipment and detection algorithm, the detection error is large, which affects the small target identification and track prediction of the sea scene, and further becomes the bottleneck of many practical applications. Therefore, the detection error of the small sea target is large, and the problem of affecting the small target identification of the sea scene is urgent to be solved. SUMMARY

[0004] The main purpose of the present application is to provide a sea target identification method and device, which solves the problem of large error of the existing small sea target detection method and affects the accuracy of small target identification in the sea scene.

[0005] In order to achieve the above purpose, according to the first aspect of the present application, a sea target identification method is provided.

[0006] The method for identifying a marine target according to the present application comprises: performing target detection on a plurality of signal sources of sensors to obtain a plurality of detection results, each detection result being a detection result of a candidate target set; recording the plurality of detection results in a unified preset form, the recorded content including at least coordinates, size, and confidence; aligning the coordinate systems of the plurality of sensors to obtain a new scene representation space according to the calibration of the plurality of sensors; projecting the plurality of detection results into the new scene representation space to obtain a plurality of new detection results corresponding to the plurality of detection results, and recording the plurality of new detection results in the unified preset form; fusing the plurality of new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set; and identifying the target type of each target according to a multi-data fusion target identification algorithm, the multi-data fusion target identification algorithm being obtained by training sample data in the new scene representation space.

[0007] Optionally, the fusing of the plurality of new detection results according to the preset fusion algorithm comprises: performing unique fusion calculation on the plurality of new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target; and performing data updating on the fusion detection result corresponding to each target, the data updating including size updating, coordinate updating, and confidence updating.

[0008] Optionally, the performing of unique fusion calculation on the plurality of new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target comprises: performing similarity calculation on all new detection results in the plurality of new detection results; judging whether different new detection results belong to the same target according to the similarity calculation result; if the different new detection results belong to the same target, performing unique fusion on the new detection results belonging to the same target to obtain the fusion detection result corresponding to the target; and if the different new detection results do not belong to the same target, the different new detection results are not fused, and the new detection results are taken as the fusion detection result corresponding to the target.

[0009] Optionally, the performing of data updating on the fusion detection result corresponding to each target comprises: if the fusion detection result is not obtained through unique fusion, the confidence of the fusion detection result is the confidence of the corresponding new detection result; if the fusion detection result is obtained through unique fusion, and the new detection results subjected to unique fusion are detection results of the same type, the confidence of the fusion detection result is the mean of the confidences of the new detection results subjected to unique fusion; and if the fusion detection result is obtained through unique fusion, and the new detection results subjected to unique fusion are detection results of different types, the confidence of the fusion detection result is an enhanced calculation result of the confidences of the new detection results subjected to unique fusion.

[0010] Optionally, the method further comprises constructing the multi-data fusion target recognition algorithm, and the constructing the multi-data fusion target recognition algorithm comprises: collecting a preset number of targets and labeling target types of the preset number of targets; representing the collected preset number of targets in the new scene representation space; generating the sample data from the preset number of targets represented in the new scene representation space and the corresponding target types; and training a deep learning model based on the sample data to obtain the multi-data fusion target recognition algorithm.

[0011] Optionally, before the plurality of new detection results are fused according to the preset fusion algorithm, the method further comprises: deleting a new detection result with a confidence less than a preset confidence threshold from the plurality of new detection results.

[0012] Optionally, the similarity calculation on all new detection results in the plurality of new detection results comprises: a pairwise intersection over union calculation on all new detection results in the plurality of new detection results; and the judging whether different new detection results belong to the same target according to the similarity calculation result comprises: judging whether different new detection results belong to the same target according to a comparison result of the intersection over union calculation result and a preset similarity threshold.

[0013] Optionally, the pairwise intersection over union calculation on all new detection results in the plurality of new detection results comprises: a pairwise intersection over union calculation on sizes in all new detection results in the plurality of new detection results, and the size is an area of a target or a volume of the target.

[0014] Optionally, the unique fusion of the new detection results belonging to the same target to obtain the fusion detection result corresponding to the target comprises: taking a largest size in the new detection results belonging to the same target as the fusion detection result corresponding to the target.

[0015] In order to achieve the above-mentioned purpose, according to the second aspect of the present application, a sea target recognition device is provided.

[0016] The marine target recognition device according to the present application comprises: a detection unit configured to perform target detection on signal sources of a plurality of sensors respectively to obtain a plurality of detection results, each detection result being a detection result of a candidate target set; a recording unit configured to record the plurality of detection results in a unified preset form, the recorded content including at least coordinates, size, and confidence; a generation unit configured to align coordinate systems of the plurality of sensors according to calibration of the plurality of sensors to obtain a new scene representation space; a projection unit configured to project the plurality of detection results into the new scene representation space respectively to obtain a plurality of new detection results, and record the plurality of new detection results in the unified preset form; a fusion unit configured to fuse the plurality of new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set; and a recognition unit configured to perform target recognition on all the fusion detection results according to a multi-data fusion target recognition algorithm to determine a target type of each target, the multi-data fusion target recognition algorithm being obtained by training sample data in the new scene representation space.

[0017] Optionally, the fusion unit comprises: a calculation module configured to perform unique fusion calculation on the plurality of new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target; and an update module configured to perform data update on the fusion detection result corresponding to each target, the data update including size update, coordinate update, and confidence update.

[0018] Optionally, the calculation module is configured to: perform similarity calculation on all the new detection results in the plurality of new detection results; determine whether different new detection results belong to the same target according to the similarity calculation result; if the different new detection results belong to the same target, perform unique fusion on the new detection results belonging to the same target to obtain the fusion detection result corresponding to the target; and if the different new detection results do not belong to the same target, do not fuse the different new detection results, and take the new detection results as the fusion detection result corresponding to the target.

[0019] Optionally, the update module is configured to: if the fusion detection result is not obtained through unique fusion, the confidence of the fusion detection result is the confidence of the corresponding new detection result; if the fusion detection result is obtained through unique fusion, and the new detection results performing the unique fusion are detection results of the same type, the confidence of the fusion detection result is the average of the confidences of the new detection results performing the unique fusion; and if the fusion detection result is obtained through unique fusion, and the new detection results performing the unique fusion are detection results of different types, the confidence of the fusion detection result is an enhanced calculation result of the confidences of the new detection results performing the unique fusion.

[0020] Optionally, the apparatus further comprises a constructing unit configured to construct the multi-data fusion target recognition algorithm, the constructing unit comprising: an acquisition module configured to acquire a preset number of targets and label the preset number of targets with target types; an expression module configured to express the acquired preset number of targets in the new scene representation space; a generation module configured to generate the sample data from the preset number of targets expressed in the new scene representation space and the corresponding target types; and a training module configured to train the deep learning model based on the sample data to obtain the multi-data fusion target recognition algorithm.

[0021] Optionally, the apparatus further comprises a deleting unit configured to delete a new detection result with a confidence less than a preset confidence threshold from the plurality of new detection results before fusing the plurality of new detection results according to the preset fusion algorithm.

[0022] Optionally, the computing module is further configured to: perform pairwise intersection over union calculation on all detection results in the plurality of new detection results; and determine whether different new detection results belong to the same target according to a comparison result of the intersection over union calculation result and a preset similarity threshold.

[0023] Optionally, the computing module is further configured to: perform pairwise intersection over union calculation on sizes in all detection results in the plurality of new detection results, the size being an area of a target or a volume of a target.

[0024] Optionally, the computing module is further configured to: take a new detection result with the largest size as a fusion detection result corresponding to the same target.

[0025] To achieve the above object, according to a third aspect of the present application, a computer readable storage medium is provided, the computer readable storage medium stores computer instructions, the computer instructions are used to make the computer execute the sea target recognition method in any one of the first aspect.

[0026] To achieve the above object, according to a fourth aspect of the present application, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores computer programs executable by the at least one processor, and the computer programs are executed by the at least one processor to make the at least one processor execute the sea target recognition method in any one of the first aspect.

[0027] In the method and device for identifying a sea target according to the embodiments of the present application, the detection results of multiple sensors are projected into a new scene representation space to obtain new detection results, so that the multiple detection results are aligned, and then the new detection results are fused uniquely to obtain more reasonable detection results with higher confidence, thereby reducing the detection errors caused by defects of the detection equipment and the detection algorithm. In addition, the target recognition algorithm for multiple data fusion is trained in the new scene representation space to identify the target, so that the accuracy of the algorithm for identifying small targets is improved, and the accuracy of the sea small target identification is improved as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and to make apparent the other features and advantages of the present application. The illustrative embodiments of the present application are described in connection with the accompanying drawings, of which:

[0029] Figure 1 is a flow chart of a method for identifying a sea target according to an embodiment of the present application;

[0030] Figure 2 is a block diagram of a device for identifying a sea target according to an embodiment of the present application;

[0031] Figure 3 is a block diagram of another device for identifying a sea target according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the technical personnel in the art better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0033] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described accompanying drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or device.

[0034] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict.

[0035] The inventor finds that, in the sea target detection stage, different detection devices adopt different signal sources, obtain different forms of detection data, and adopt different detection algorithms, so that the error of the detection result in the detection stage is mainly affected by the system error of the signal source and the detection device itself and the detection accuracy of the algorithm; the error in the sea target recognition stage is mainly the calculation error from the related algorithm itself, and in addition, the error in the target detection stage also accumulates to the stage to affect the recognition result. The sea target recognition method in the embodiments of the present application mainly aligns and projects the detection result to a new scene representation space based on a data alignment fusion method, so as to reduce the error in the target detection stage, and then performs target recognition in the new scene representation space based on the foregoing target detection result, reduces the recognition error, and further improves the accuracy of target recognition. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0036] According to the embodiments of the present application, a sea target recognition method is provided, as shown in Figure 1 The method comprises the following steps S101-S106: S101. performing target detection on the signal sources of multiple sensors to obtain multiple detection results; wherein each detection result is the detection result of a candidate target set; S102. recording the multiple detection results in a unified preset form; wherein the recorded content at least includes coordinates, size, and confidence; S103. aligning the coordinate systems of the multiple sensors according to the calibration of the multiple sensors to obtain a new scene representation space; S104. projecting the multiple detection results into the new scene representation space to obtain corresponding multiple new detection results, and recording the multiple new detection results in a unified preset form; S105. fusing the multiple new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set; S106. performing target recognition on all the fusion detection results according to a multi-data fusion target recognition algorithm to determine the target type of each target; wherein the multi-data fusion target recognition algorithm is obtained by training sample data in the new scene representation space.

[0037] In step S101, the plurality of sensors can be infrared sensors, optical sensors, electromagnetic radar sensors, synthetic aperture radar sensors, etc. In actual applications, an offshore identification or tracking system generally carries two sensors for target detection in a marine scene. The embodiments of the present application do not limit the types of sensors, and there can be more than two sensors. For the signals returned by each sensor, this step selects a suitable detection algorithm to carry out target detection. For example, for image signals returned by an optical sensor, a multi-target detection algorithm based on computer vision can be used, such as YOLO5, DETR, etc.

[0038] Each detection result is a detection result of a candidate target set, which means that each detection result contains a plurality of offshore targets, which form a candidate target set. In subsequent target tracking, the target of interest is selected from the candidate target set for tracking analysis according to actual business requirements. One sensor corresponds to one detection result. If there are multiple sensors of the same type, one sensor corresponds to one detection result.

[0039] In step S102, in order to facilitate subsequent fusion, each detection result obtained in step S101 is stored and recorded in a unified form (preset form). The recorded content at least includes coordinates (such as X and Y values under two-dimensional coordinates; or X, Y, and Z values under three-dimensional coordinates), size, confidence, and can also include other information commonly used in the field of target detection. The coordinates and size recorded in this step are coordinates and size under the corresponding sensor coordinates.

[0040] In step S103, coordinate system alignment means that the coordinate systems of the plurality of sensors are transformed to obtain the same coordinate system (new scene representation space). The coordinate system alignment is to convert the targets represented in different coordinate systems to the same coordinate system for representation, and to realize the alignment of multiple data. It should be noted that the form of the new scene representation space in the embodiments of the present application is not limited, but must be a representation space with stronger abstraction ability.

[0041] Specifically, it is assumed that the coordinate system corresponding to the infrared sensor is Γ, the coordinate system corresponding to the optical sensor is I, the new scene representation space is Φ, and the mapping functions of Γ and I projected to Φ are P r and P i For any target O, the following logic is true:

[0042] P r (O Γ )≡P i (O I )

[0043] Where O Γ represents the real detection result of O in the coordinate system Γ corresponding to the infrared sensor, and OI represents the real detection result of O in the coordinate system I corresponding to the optical sensor, and ≡ represents that each data (such as coordinates, size, etc.) of the detection result is equal. It should be noted that the real detection result here is a concept used to explain the use of multi-data alignment, and is not a detection result obtained by a detection algorithm, but refers to an objective fact in the real world. It should be noted that the mapping function can be obtained based on the reverse calculation of the coordinate data before and after the coordinate conversion.

[0044] In step S104, "projecting the plurality of detection results into the new scene representation space to obtain a plurality of new detection results corresponding to the new scene representation space" means that each detection result in the plurality of detection results is projected into the new scene representation space to obtain a new detection result based on different mapping functions corresponding to different sensors (the mapping function used to project into the new scene representation space in step S103). The new detection result also needs to be recorded in the unified preset form as described above, and the recording content at least includes coordinates, size, confidence, and can also include other information commonly used in target detection. The coordinates and size corresponding to the new detection result are the coordinates and size in the new scene representation space, and the confidence is the confidence of the detection result in step S101.

[0045] In step S105, the plurality of new detection results are fused according to the preset fusion algorithm, which includes:

[0046] First, the plurality of new detection results are uniquely fused according to the preset fusion algorithm to obtain a fusion detection result corresponding to each target; then, the fusion detection result corresponding to each target is updated, and the data update includes size update, coordinate update and confidence update.

[0047] Specifically, the plurality of new detection results are uniquely fused according to the preset fusion algorithm to obtain a fusion detection result corresponding to each target, which is specifically implemented as follows: 1) similarity calculation is performed on all new detection results in the plurality of new detection results; 2) whether different new detection results belong to the same target is determined according to the similarity calculation result; 3) if they belong to the same target, the new detection results belonging to the same target are uniquely fused to obtain a fusion detection result corresponding to the target; and 4) if they do not belong to the same target, different new detection results are not fused, and the new detection result is taken as the fusion detection result corresponding to the target.

[0048] The assumption basis of unique fusion is that in the same scene at the same time, each target has and only has one corresponding detection result.

[0049] The application embodiment gives a similarity calculation method: pairwise intersection-over-union calculation is performed on all detection results in the plurality of new detection results, and the following is the intersection-over-union calculation formula:

[0050]

[0051] Wherein, IOU is the intersection over union of two new detection results, A and B can represent the size of two new detection results respectively, and the size can be area or volume. If the new scene representation space is two-dimensional space, A and B represent the area of two new detection results respectively, and if the new scene representation space is three-dimensional space, A and B represent the volume of two new detection results respectively.

[0052] In the above example, IOU is the similarity calculation result, and "determining whether different new detection results belong to the same target according to the similarity calculation result" means determining whether it belongs to the same target according to the comparison result of the intersection over union calculation result IOU and the preset similarity threshold θ. Specifically, if IOU > θ, it is determined that different new detection results belong to the same target; if IOU ≤ θ, it is determined that different new detection results do not belong to the same target. It should be noted that for the case where the number of sensor types or the number of sensors is greater than or equal to 3, through similarity calculation, more than or equal to 3 new detection results belonging to the same target can be finally obtained.

[0053] "Uniquely fusing new detection results belonging to the same target to obtain a fusion detection result corresponding to the target" can be to take the largest size of the new detection results belonging to the same target as the fusion detection result corresponding to the target. Of course, it can also be fused in other ways, such as taking the mean value of multiple new detection results, taking the mean value result as the fusion detection result corresponding to the target; or selecting the highest confidence as the fusion detection result corresponding to the target.

[0054] Specifically, the update of the confidence in the data update of each target corresponding fusion detection result is: if the fusion detection result is not obtained through unique fusion, the confidence of the fusion detection result is the confidence of the corresponding new detection result; if the fusion detection result is obtained through unique fusion, and the new detection results for unique fusion are detection results of the same type, then the confidence of the fusion detection result is the mean value of the confidence of the new detection results for unique fusion; if the fusion detection result is obtained through unique fusion, and the new detection results for unique fusion are detection results of different types, then the confidence of the fusion detection result is the enhanced calculation result of the confidence of the new detection results for unique fusion.

[0055] The confidence calculation for the case where the fusion detection result is obtained through unique fusion is further illustrated by the following formula:

[0056]

[0057] Wherein, conf 同 And conf 异respectively represent the confidence of the unique fusion of the same type of detection result and the different type of detection result, n and m respectively represent the number of new detection results participating in the unique fusion, n can be set by a manual threshold, and the effect of n-neighbor fusion is achieved, and m is the maximum value of the number of sensors for target detection, conf i and conf j respectively represent the new detection results participating in the unique fusion, # represents the number operator for counting the number of elements in the set pointed to by i and j. It should be noted that for the case of "taking the largest size among the new detection results belonging to the same target as the fusion detection result corresponding to the target", conf j is the confidence of the new detection result with the largest size among the different types of detection results belonging to the same target.

[0058] The foregoing is the update of the confidence, and the update of the size and the coordinate means adding the corresponding size and coordinate to each fusion detection result.

[0059] In step S106, the multi-data fusion target recognition algorithm is constructed in advance. The construction of the multi-data fusion target recognition algorithm includes: collecting a preset number of targets, and labeling the target types of the preset number of targets; representing the collected preset number of targets in a new scene representation space; generating sample data from the preset number of targets represented in the new scene representation space and the corresponding target types; and training a deep learning model based on the sample data to obtain the multi-data fusion target recognition algorithm. The target in this step is a small target at sea, and the target recognition method in the embodiments of the present application is mainly for small targets at sea. The definition of small targets is based on the mainstream standards in academia and industry. In addition, the labeled target types include canoes, fishing boats, warships, or fish, etc. The preset number of targets are detected by multiple sensors to obtain detection results, and the target types of the preset number of targets are labeled, which means that the corresponding target type is labeled for each detection result.

[0060] "Performing target recognition on all fusion detection results according to the multi-data fusion target recognition algorithm to determine the target type of each target" means inputting all fusion detection results obtained in step S105 into the multi-data fusion target recognition algorithm, and outputting the target types corresponding to all fusion detection results.

[0061] From the above description, it can be seen that in the sea target identification method of the embodiment of the application, the multiple detection results corresponding to multiple sensors are projected into a new scene representation space to obtain new detection results, alignment of the multiple detection results is realized, and then the new detection results are uniquely fused to obtain more reasonable detection results with higher confidence, thereby reducing detection errors caused by defects of detection equipment and detection algorithms. In addition, the target identification algorithm for multi-data fusion is trained in the new scene representation space to perform target identification, the accuracy of the algorithm in identifying small targets is improved, and thus the accuracy of sea small target identification is improved as a whole.

[0062] Further, in order to better describe the relative degree of confidence, after the confidence of all the fused detection results in step S105 is calculated, normalization calculation can also be performed on the fused detection results, and the calculation method is as follows:

[0063]

[0064] wherein, indicates the confidence of the normalized fused detection result, K indicates the number of all the fused detection results, and respectively indicate maximum value and minimum value operations.

[0065] Further, the multi-data fusion target identification algorithm in step S106 can also be any algorithm based on the classification idea, such as the nearest neighbor algorithm. On the basis of having a sample library, the nearest neighbor calculation is performed between the new detection result to be identified and the sample data with a label, and the most similar class label can be obtained, that is, the target type of the new detection result.

[0066] Further, in order to improve the accuracy of identification, before the multiple new detection results are fused according to the preset fusion algorithm, the method further includes: deleting the new detection result with a confidence less than a preset confidence threshold from the multiple new detection results. The new detection result with a confidence less than a preset confidence threshold is deleted from the multiple new detection results in order to remove redundant detection results. The preset confidence threshold can be manually set; or the preset confidence threshold can be an adaptive threshold. The adaptive threshold can be determined in multiple forms, such as an adaptive threshold selection method based on statistical sorting.

[0067] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0068] According to the embodiment of the application, a computer readable storage medium is also provided, which stores the computer executable instructions. Figure 1The method for identifying a marine target 200, as shown in Figure 2 The device includes: a detection unit 21 for detecting a plurality of sensor signal sources to obtain a plurality of detection results, each detection result being a detection result of a candidate target set; a recording unit 22 for recording the plurality of detection results in a unified preset form, the recorded content including at least coordinates, size, and confidence; a generation unit 23 for aligning the coordinate systems of the plurality of sensors to obtain a new scene representation space according to the calibration of the plurality of sensors; a projection unit 24 for projecting the plurality of detection results into the new scene representation space to obtain corresponding new detection results, and recording the plurality of new detection results in a unified preset form; a fusion unit 25 for fusing the plurality of new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set; and an identification unit 26 for identifying the target type of each target according to a multi-data fusion target identification algorithm trained on sample data in the new scene representation space.

[0069] Further, as shown in Figure 3 The fusion unit 25 includes: a calculation module 251 for performing unique fusion calculation on the plurality of new detection results to obtain a fusion detection result corresponding to each target; and an update module 252 for updating the fusion detection result corresponding to each target, the data update including size update, coordinate update, and confidence update.

[0070] Further, as shown in Figure 3 The calculation module 251 is configured to: perform similarity calculation on all new detection results in the plurality of new detection results; determine whether different new detection results belong to the same target according to the similarity calculation result; if they belong to the same target, perform unique fusion on the new detection results belonging to the same target to obtain a fusion detection result corresponding to the target; and if they do not belong to the same target, do not fuse the different new detection results, and take the new detection results as the fusion detection result corresponding to the target.

[0071] Further, as shown in Figure 3As shown, the updating module 252 is configured to: if the fusion detection result is not obtained through the uniqueness fusion, the confidence of the fusion detection result is the confidence of the corresponding new detection result; if the fusion detection result is obtained through the uniqueness fusion, and the new detection result for the uniqueness fusion is the same type of detection result, the confidence of the fusion detection result is the average of the confidence of the new detection result for the uniqueness fusion; if the fusion detection result is obtained through the uniqueness fusion, and the new detection result for the uniqueness fusion is the different type of detection result, the confidence of the fusion detection result is the enhanced calculation result of the confidence of the new detection result for the uniqueness fusion.

[0072] Further, as shown in Figure 3 The apparatus further includes a constructing unit 27 configured to construct the multi-data fusion target recognition algorithm, the constructing unit 27 includes: an acquisition module 271 configured to acquire a preset number of targets and label the target types of the preset number of targets; an expression module 272 configured to express the acquired preset number of targets in a new scene expression space; a generation module 273 configured to generate sample data from the preset number of targets expressed in the new scene expression space and the corresponding target types; and a training module 274 configured to train the deep learning model based on the sample data to obtain the multi-data fusion target recognition algorithm.

[0073] Further, as shown in Figure 3 The apparatus further includes a deleting unit 28 configured to delete a new detection result with a confidence less than a preset confidence threshold from the plurality of new detection results before the plurality of new detection results are fused according to the preset fusion algorithm.

[0074] Further, as shown in Figure 3 The calculating module 251 is further configured to: perform pairwise intersection over union calculation on all detection results in the plurality of new detection results; and determine whether different new detection results belong to the same target according to a comparison result of the intersection over union calculation result and a preset similarity threshold.

[0075] Further, as shown in Figure 3 The calculating module 251 is further configured to: perform pairwise intersection over union calculation on the sizes of all detection results in the plurality of new detection results, the size being an area of the target or a volume of the target.

[0076] Further, as shown in Figure 3 The calculating module 251 is further configured to: take the new detection result with the largest size as the fusion detection result corresponding to the target.

[0077] Specifically, the specific process in which each unit and module in the apparatus of the embodiment of the present application realizes its function can be referred to the related description in the method embodiment, which will not be repeated here.

[0078] From the above description, it can be seen that in the sea target identification device of the embodiment of the application, the multiple detection results corresponding to multiple sensors are projected into a new scene representation space to obtain new detection results, alignment of the multiple detection results is realized, and then the new detection results are uniquely fused to obtain more reasonable and higher confidence detection results, so that detection errors caused by defects of detection equipment and detection algorithms are reduced. In addition, the target identification algorithm for multi-data fusion is trained in the new scene representation space to perform target identification, the precision of the algorithm for identifying small targets is improved, and then the accuracy of sea small target identification is improved as a whole.

[0079] According to the embodiment of the application, a computer readable storage medium is also provided, characterized in that the computer readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the sea target identification method in the method embodiment.

[0080] According to the embodiment of the application, an electronic device is also provided, including at least one processor, and a memory in communication connection with the at least one processor; wherein the memory stores computer programs executable by the at least one processor, and the computer programs are executed by the at least one processor to make the at least one processor execute the sea target identification method in the method embodiment.

[0081] Obviously, those skilled in the art should understand that the modules or steps of the application described above can be realized by general computing devices, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by computing devices, so that they can be stored in storage devices and executed by computing devices, or they can be respectively manufactured into individual integrated circuit modules, or multiple modules or steps among them can be manufactured into a single integrated circuit module to realize. Thus, the application is not limited to any specific combination of hardware and software.

[0082] The above only describes the preferred embodiments of the application and is not used to limit the application. For those skilled in the art, the application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method of identification of a maritime target, characterized in that, The method comprises: target detection is performed on a plurality of sensor signal sources respectively to obtain a plurality of detection results, each detection result being a detection result of a candidate target set; the plurality of detection results are recorded in a unified preset form, and the recorded content at least includes coordinates, sizes, and confidence levels; a new scene representation space is obtained by aligning the coordinate systems of the plurality of sensors according to the calibration of the plurality of sensors; the plurality of detection results are respectively projected into the new scene representation space to obtain corresponding new detection results, and the plurality of new detection results are recorded in the unified preset form; the plurality of new detection results are fused according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set; the fusion of the plurality of new detection results according to the preset fusion algorithm comprises: performing unique fusion calculation on the plurality of new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target; and performing data updating on the fusion detection result corresponding to each target, wherein the data updating comprises size updating, coordinate updating, and confidence level updating; the unique fusion calculation on the plurality of new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target comprises: performing similarity calculation on all new detection results in the plurality of new detection results; judging whether different new detection results belong to the same target according to the similarity calculation result; if the new detection results belong to the same target, performing unique fusion on the new detection results belonging to the same target to obtain the fusion detection result corresponding to the target; and if the new detection results do not belong to the same target, the different new detection results are not fused, and the new detection results are taken as the fusion detection result corresponding to the target; the data updating on the fusion detection result corresponding to each target comprises: if the fusion detection result is not obtained through unique fusion, the confidence level of the fusion detection result is the confidence level of the corresponding new detection result; if the fusion detection result is obtained through unique fusion, and the new detection results subjected to unique fusion are detection results of the same type, the confidence level of the fusion detection result is the average of the confidence levels of the new detection results subjected to unique fusion; and if the fusion detection result is obtained through unique fusion, and the new detection results subjected to unique fusion are detection results of different types, the confidence level of the fusion detection result is an enhanced calculation result of the confidence levels of the new detection results subjected to unique fusion; target recognition is performed on all fusion detection results according to a multi-data fusion target recognition algorithm to determine the target type of each target, wherein the multi-data fusion target recognition algorithm is obtained by training sample data in the new scene representation space.

2. The method of identifying a maritime target according to claim 1, characterized in that, The method further comprises constructing the multi-data fusion target recognition algorithm, and the construction of the multi-data fusion target recognition algorithm comprises: a preset number of targets are collected, and the target types of the preset number of targets are labeled; the collected preset number of targets are represented in the new scene representation space; and a preset number of targets represented in the new scene representation space and corresponding target types are generated as the sample data; a deep learning model is trained based on the sample data to obtain the multi-data fusion target recognition algorithm.

3. The method of identifying a maritime target according to claim 1, characterized in that, Before the plurality of new detection results are fused according to the preset fusion algorithm, the method further comprises: deleting a new detection result from the plurality of new detection results, wherein the confidence of the new detection result is less than a preset confidence threshold.

4. The method of claim 1, wherein, The similarity calculation on all new detection results in the plurality of new detection results comprises: a pair-wise intersection over union calculation is performed on all new detection results in the plurality of new detection results. The determination of whether different new detection results belong to the same target according to the similarity calculation result comprises: a determination of whether different new detection results belong to the same target according to a comparison result of the intersection over union calculation result and a preset similarity threshold.

5. A device for identification of a marine target, characterized in that The device comprises: a detection unit configured to perform target detection on a plurality of signal sources of sensors respectively to obtain a plurality of detection results, each detection result being a detection result of a candidate target set; a recording unit configured to record the plurality of detection results in a unified preset form, the recorded content at least including coordinates, sizes, and confidences; a generation unit configured to align coordinate systems of the plurality of sensors to obtain a new scene representation space according to calibration of the plurality of sensors; a projection unit configured to project the plurality of detection results into the new scene representation space respectively to obtain corresponding plurality of new detection results, and record the plurality of new detection results in the unified preset form; a fusion unit configured to fuse the plurality of new detection results according to a preset fusion algorithm to obtain a fusion detection result corresponding to each target in the candidate target set. The fusing of the multiple new detection results according to the preset fusion algorithm comprises: performing unique fusion calculation on the multiple new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target; performing data updating on the fusion detection result corresponding to each target, wherein the data updating comprises size updating, coordinate updating and confidence updating; the unique fusion calculation on the multiple new detection results according to the preset fusion algorithm to obtain the fusion detection result corresponding to each target comprises: performing similarity calculation on all new detection results in the multiple new detection results; judging whether different new detection results belong to the same target according to the similarity calculation result; if the new detection results belong to the same target, performing unique fusion on the new detection results belonging to the same target to obtain the fusion detection result corresponding to the target; if the new detection results do not belong to the same target, the different new detection results are not fused, and the new detection results are taken as the fusion detection result corresponding to the target; the data updating on the fusion detection result corresponding to each target comprises: if the fusion detection result is not obtained through unique fusion, the confidence of the fusion detection result is the confidence of the corresponding new detection result; if the fusion detection result is obtained through unique fusion, and the new detection results performing unique fusion are detection results of the same type, the confidence of the fusion detection result is the mean of the confidences of the new detection results performing unique fusion; if the fusion detection result is obtained through unique fusion, and the new detection results performing unique fusion are detection results of different types, the confidence of the fusion detection result is the enhanced calculation result of the confidences of the new detection results performing unique fusion. The recognition unit is configured to perform target recognition on all fusion detection results according to a multi-data fusion target recognition algorithm to determine the target type of each target, wherein the multi-data fusion target recognition algorithm is obtained by training sample data in the new scene representation space.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing the computer to execute the method for identifying a marine target according to any one of claims 1 to 4.

7. An electronic device, comprising: The method comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to cause the at least one processor to execute the method for identifying a marine target according to any one of claims 1 to 4.

Images