Aerial motion target dual detection method and system, and storage medium

Abstract

The application provides a kind of aerial sports target dual detection method and system, storage medium, its method steps include: establishing cyclically updated global dynamic event group E, the polygon modeling of motion target in scene label graph is carried out, constructs space- time neighborhood, to extract out local dynamic event group F in E;When judging the density of F meets the standard, inflate space-time neighborhood morphs to cube to be used as candidate target area;Candidate target area is mapped into binary image G;According to the time and position information of dynamic event in candidate target area, a new label graph is intercepted in the corresponding scene label graph;Establish neural network, whether it is motion target according to new label graph and binary image G is judged;Global target recognition network is established, all dynamic events in candidate target area are used as label, E is input, training is completed to output motion target information. In this way, the problem that bionic dynamic event cannot be used for neural network training and detection is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of moving target detection, in particular to a dual detection method and system for aerial moving targets based on a bionic dynamic event and a multi-edge positioning scheme, and a storage medium. BACKGROUND

[0002] Due to the huge amount of video sequence data obtained by traditional visual sensors and the high information redundancy, these shortcomings result in slow running speed and poor real-time performance of target detection algorithms based on such videos.

[0003] The dynamic events output by the bionic dynamic sensor (DVS / EVS) have strong complementary advantages in sparsity, anti-motion blur, dynamic range, and requirements for light, compared with videos obtained by traditional visual sensors. However, the problem is that when the bionic dynamic event is applied to the detection of moving objects in aerial photography, the dynamic signal contains less information and cannot be directly processed by existing target detection algorithms.

[0004] On the other hand, the existing target detection algorithms based on traditional aerial videos with the help of various complex neural network models have greatly improved the detection accuracy compared with previous algorithms. However, for high-resolution images provided by unmanned aerial vehicles, these algorithms still need to rely on sliding window technology to cover the entire image, so under the limitation of processor computing power, these algorithms run slowly and are difficult to be directly applied to low-power image processing platforms such as unmanned aerial vehicles. Even if a lightweight network model is used in the detection module, it is difficult to achieve a good balance between detection accuracy and detection real-time performance.

[0005] In addition, the existing target detection algorithms based on bionic dynamic events mostly integrate dynamic events within a certain time interval to form static frame images in order to directly use deep learning algorithms for research. However, this processing method can integrate bionic dynamic events and deep learning into the target detection framework, but it cannot fully utilize the sparsity and asynchrony of bionic dynamic events, and the static frame images formed by integration cannot express the time information contained in the bionic dynamic events. Therefore, these problems result in the inability of existing technologies to use bionic dynamic events for neural network training and detection. SUMMARY

[0006] Therefore, the main purpose of the present application is to provide a dual detection method and system for aerial moving targets, and a storage medium, to realize direct processing of asynchronous and sparse bionic dynamic events to focus on the area with candidate moving targets, thereby solving the problem that bionic dynamic events cannot be used for neural network training and detection.

[0007] In order to achieve the above purpose, according to a first aspect of the present application, a dual detection method for aerial moving targets is provided, which comprises the following steps:

[0008] Step S100 obtains the bionic dynamic event stream, establishes a cyclically updated global dynamic event group E, polygonally models the moving target in the scene label graph, constructs a spatio-temporal neighborhood, and extracts a local dynamic event group F in E;

[0009] Step S200 inflates the spatio-temporal neighborhood to a cube as a candidate target region when judging that the density of the local dynamic event group F meets the standard;

[0010] Step S300 maps the candidate target region into a binary image G; and according to the time and position information of the dynamic events in the candidate target region, a new label graph is intercepted in the corresponding scene label graph;

[0011] Step S400 establishes a target contour recognition network to extract the contour of the moving target in the new label graph; and establishes a target detection network, takes the binary image G and the contour information of the moving target as input, and performs fusion training to determine whether the binary image G is a moving target;

[0012] Step S500 establishes a global target recognition network, takes all the dynamic events in the candidate target region as labels, takes the global dynamic event group E as input, completes training, and outputs the moving target information.

[0013] In a possible preferred embodiment, in step S100, the step of establishing a cyclically updated global dynamic event group E includes:

[0014] Step S110 buffers the bionic dynamic event stream in a time period to establish the global dynamic event group E;

[0015] Step S120 reads a new dynamic event e from the bionic dynamic event stream, finds the minimum value of the time of all the dynamic events in the global dynamic event group E , and takes the time information t of e as a reference. > If the absolute value of the difference between t and t is greater than , the dynamic events in the global dynamic event group E that have a time stamp different from t by more than need to be deleted to refresh the global dynamic event group E.

[0016] In a possible preferred embodiment, in step S100, the step of polygonal modeling includes:

[0017] Step S130 selects a single moving target from the scene label graph, constructs a circumscribed circle with the length of the moving target as the diameter and the center coordinate of the moving target as the center of the circle.

[0018] Step S140 constructs a polygon in the circumcircle to enclose the moving target; calculates the distance and angle of each vertex of the polygon relative to the center coordinates of the moving target to construct a corresponding relation function , and completes the modeling.

[0019] In a possible preferred embodiment, in step S100, the spatiotemporal neighborhood constructing step comprises:

[0020] Step S150 constructs a polygon centered at the position information of the dynamic event e as the spatial neighborhood according to the relation function

[0021] Step S160 selects as the length to stretch the polygon spatial neighborhood to construct the spatiotemporal neighborhood.

[0022] In a possible preferred embodiment, in step S200, the step further comprises:

[0023] Step S210, when judging that the density of the local dynamic event group F does not meet the requirement, adjusts the spatial position of the spatiotemporal neighborhood within a preset number of times until the requirement is met, and executes step S300.

[0024] Step S220, if the requirement is still not met, stores the current dynamic event e in the global dynamic event group E, and reads a new dynamic event e to execute step S200.

[0025] In a possible preferred embodiment, in step S200, the step further comprises:

[0026] Step S230 deletes all dynamic events in the local dynamic event group F judged as the candidate target region from the global dynamic event group E to update the global dynamic event group E.

[0027] In a possible preferred embodiment, in step S300, the step of mapping the candidate target region into a binary image G comprises:

[0028] Step S310 initializes a binary image P with a resolution of L L, and sets the pixel value of each position to 0, wherein L is the length of the square candidate target region;

[0029] Step S320 reads all dynamic events in the local dynamic event group F of the candidate target region, and maps each dynamic event into the binary image P according to the position information of the dynamic event;

[0030] Step S330 sets the pixel value of the position mapped by all dynamic events in P to 255 to obtain the binary image G.

[0031] ​To achieve the above object, corresponding to the above method, the second aspect of the present application also provides a dual detection system for aerial sports targets, which comprises:

[0032] A storage unit is configured to store a program comprising steps of the dual detection method for aerial sports targets as any one of the above methods, so as to be executed by the positioning unit, the first detection unit, the training set processing unit, the second detection unit, and the timely calling unit.

[0033] The positioning unit is configured to buffer a bionic dynamic event stream in a time period, establish a global dynamic event group E, and cyclically update; construct a spatial neighborhood with a newly read dynamic event e as the center and a polygon modeling of the moving target as the space neighborhood, a spatiotemporal neighborhood for length, so as to extract a local dynamic event group F in E; judge the density of the local dynamic event group F, when the density reaches a standard, inflate the spatiotemporal neighborhood to a cube to serve as a candidate target region, and delete the local dynamic event group F in the candidate target region from the global dynamic event group E;

[0034] The training set processing unit is configured to map the candidate target region into a binary image G according to coordinate information; find a scene label image corresponding to a time stamp according to time information of e in the candidate target region, and demarcate the corresponding candidate target region in the scene label image according to the position information of e, and intercept a new label image.

[0035] The first detection unit is configured to identify whether the new label image and the binary image G are moving targets.

[0036] The second detection unit is configured to take all dynamic events in the candidate target region corresponding to the positive sample as a label, and take the global dynamic event group E as an input to complete training of a global target recognition network, and then identify moving target information in the dynamic event.

[0037] In a possible preferred embodiment, the detection unit comprises: a target contour recognition network configured to extract a moving target contour in the new label image; and a target detection network configured to take the binary image G and the moving target contour information as an input to perform fusion training, so as to judge whether the binary image G is a moving target.

[0038] To achieve the above object, corresponding to the above method, the third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement steps of the dual detection method for aerial sports targets as any one of the above methods.

[0039] The dual detection method and system for aerial motion targets, and the storage medium provided by the application can process bionic dynamic events one by one, fully utilize the asynchronous sparsity of the bionic dynamic events, and make the overall algorithm have less redundant data; meanwhile, the candidate targets are located from the bionic dynamic event stream through the innovative multi-edge modeling scheme, so that the useless information in the scene can be quickly removed, the area with the motion target is focused, and the response speed of the algorithm is improved.

[0040] In addition, the positioning result of the candidate target can solve the problem that the bionic dynamic event lacks labels when used for neural network training, so that the bionic dynamic event can be used for training and detection of the neural network, thereby fully exerting the advantages of the bionic dynamic event in the motion detection field and improving the real-time performance and detection accuracy of the overall algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this application. The embodiments of the application illustrated in the drawings, and their description, are presented to explain the application and not to limit or define the application. In the drawings:

[0042] Figure 1 A schematic diagram of the dual detection method for aerial motion targets of the application;

[0043] Figure 2 A schematic diagram of the logic framework of the dual detection method for aerial motion targets of the application;

[0044] Figure 3 A schematic diagram of the dodecagon modeling of the motion target in the scene label graph in the dual detection method for aerial motion targets of the application;

[0045] Figure 4 A schematic diagram of the dodecagon spatiotemporal neighborhood in the dual detection method for aerial motion targets of the application;

[0046] Figure 5 A schematic diagram of the extraction of the spatiotemporal neighborhood from the global dynamic event group E to the local dynamic event group F in the dual detection method for aerial motion targets of the application;

[0047] Figure 6 A schematic diagram of the inflation of the spatiotemporal neighborhood into the candidate target region of the cubic in the dual detection method for aerial motion targets of the application;

[0048] Figure 7 A schematic diagram of the mapping of the candidate target region into the binary image G according to the coordinate information in the dual detection method for aerial motion targets of the application;

[0049] Figure 8 A schematic diagram of the new label graph in the dual detection method for aerial motion targets of the application;

[0050] Figure 9 Figure 1 shows a schematic diagram of extracting a vehicle contour from a new label image in the dual detection method of the aerial sports target of the present application;

[0051] Figure 10 Figure 2 shows a schematic diagram of constructing the dual detection system of the aerial sports target of the present application. DETAILED DESCRIPTION

[0052] In order for those skilled in the art to better understand the technical solutions of the present application, the specific technical solutions of the present application will be described in detail below in conjunction with the embodiments, so as to help those skilled in the art to further understand the present application. Obviously, the embodiments described in the present application are only a part of the embodiments of the present application, not all the embodiments. It should be pointed out that, for those skilled in the art, the embodiments in the present application and the features in the embodiments can be combined with each other without departing from the concept of the present application and without being in conflict with each other. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts should belong to the disclosure and protection scope of the present application.

[0053] In addition, the terms "first", "second", "S1", "S2" and the like in the specification and claims of the present application and the drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or 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 an order other than that described herein. At the same time, the terms "include" and "have" in the present application and any variations thereof are intended to cover non-exclusive inclusion. Unless otherwise explicitly specified and limited, the terms "provide", "arrange", "mount", "connect", "connect" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood in combination with the prior art according to the specific circumstances.

[0054] In order to realize direct processing of asynchronous sparse bionic dynamic events, the region of the candidate sports target is focused, so as to solve the problem that bionic dynamic events cannot be used for neural network training and detection.

[0055] Referring to Figures 1 to 9 As shown in the figure, the present application provides a dual detection method for aerial sports target, the steps of which include:

[0056] Step S100 acquires the bionic dynamic event stream, establishes a cyclically updated global dynamic event group E, polygonally models the moving target in the scene tag graph, constructs a spatio-temporal neighborhood, and extracts a local dynamic event group F in E.

[0057] Specifically, the aerial photography device adapted by the method of the present example preferably uses an event camera and a traditional camera to coaxially photograph the same scene, simultaneously acquires the bionic dynamic signal, the scene picture, and the time information corresponding to each frame of the scene picture, and uses the scene picture as the scene tag graph of the overall algorithm.

[0058] In the present example, the step S100 of cyclically updating the global dynamic event group E includes:

[0059] Step S110 buffers the bionic dynamic event stream in a time period, and establishes the global dynamic event group E.

[0060] For example, the acquired bionic dynamic signal is sorted in chronological order to form a bionic dynamic event stream, and then events in a time period from 0 to are extracted from the sorted bionic dynamic event stream to establish the global dynamic event group E.

[0061] Step S120 reads a new dynamic event e from the bionic dynamic event stream, finds the minimum value of the time of all dynamic events in the global dynamic event group E , and takes the time information t of e as the reference. > If the absolute value of the difference between the time stamp of e and t is greater than , the dynamic event e needs to be deleted to refresh the global dynamic event group E.

[0062] Further, in step S100, the polygon modeling step includes:

[0063] Step S130 selects a single moving target from the scene tag graph, constructs an inscribed circle with the length of the moving target as the diameter and the center coordinates as the center.

[0064] Step S140 constructs a polygon in the inscribed circle to enclose the moving target, calculates the distance and angle of each vertex of the polygon relative to the center coordinates of the moving target to construct a corresponding relationship function , and completes the modeling.

[0065] Specifically, taking the moving target as a vehicle as an example, in order to more accurately extract the area where the vehicle is located, the example preferably adopts a dodecagon to model the vehicle in the scene label image, because compared with a rectangle, the dodecagon can more accurately depict the shape of the vehicle, and in addition, in other examples, a person skilled in the art can also adapt different polygons to model according to the type of the moving target and the actual shape thereof.

[0066] As shown in Figure 3 , the modeling process of the example includes: selecting a single vehicle in the scene label image, taking the length of the vehicle as a diameter to first construct an excircle of the vehicle, and the center of the excircle is the center coordinate of the vehicle . The excircle is used to preliminarily frame the area where the vehicle is located.

[0067] Secondly, a dodecagon is constructed in the excircle, wherein the upper and lower edges of the dodecagon are the chords of the excircle. Then the dodecagon is adjusted so that the vehicle can just fall within the dodecagon area, and then the distances and angles of the twelve vertices of the dodecagon relative to the center coordinate of the vehicle are calculated, and a relationship function of each vertex and the center coordinate of the vehicle is constructed .

[0068] Thus, through the relationship of the twelve vertices and the center coordinate of the vehicle , a symmetrical dodecagon capable of depicting the shape of the vehicle can be determined.

[0069] Further, in step S100, the spatiotemporal neighborhood construction step includes:

[0070] Step S150 takes the position information of the dynamic event e as the center, and constructs a polygon centered on e as a spatial neighborhood according to the relationship function .

[0071] Step S160 selects as the length in the time dimension, and stretches the polygon spatial neighborhood to construct a spatiotemporal neighborhood.

[0072] Specifically, as shown in Figures 4 to 5 , taking the position information of the dynamic event e as the center, a dodecagon centered on the dynamic event e is constructed as a spatial neighborhood according to the relationship function of the dodecagon vertex and the center point obtained in step S140 .

[0073] Then selecting as the length in the time dimension, a spatiotemporal neighborhood shaped as a dodecahedron as shown in Figure 4 can be constructed, and then all dynamic events falling within the spatiotemporal neighborhood can be extracted from the global dynamic event group E, and the extraction range is as shown in Figure 5 ​As shown, the local dynamic event group F centered on the dynamic event e is thus formed.

[0074] When the density of the local dynamic event group F is determined to be up to the standard, the spatio-temporal neighborhood is inflated to a cube to serve as a candidate target region.

[0075] Specifically, the density of the local dynamic event group F is first defined as the total number of events within it, and a threshold N is set. If the density of the local dynamic event group reaches N, it is determined to be a candidate target region, as shown in Figure 6 As shown, the dodecahedron in the x-y plane is inflated to a square, and the coordinates, time, and other information of the center point dynamic event e of the local dynamic event group F are output, and the candidate target region after the inflation is saved.

[0076] Since the local dynamic event group F at this time has been determined to be a candidate target region possibly containing a vehicle before inflation, the dynamic events entering the region due to inflation will not affect the determination of whether the region contains a vehicle. At the same time, after inflating the spatial neighborhood of the candidate target region to a square, it is more convenient for subsequent use of neural networks for further detection and processing.

[0077] On the other hand, if the density of the local dynamic event group F does not reach the threshold N, the following steps can be performed:

[0078] When the density of the local dynamic event group F is determined to be up to the standard, the spatio-temporal neighborhood is inflated to a cube to serve as a candidate target region.

[0079] If the density still does not reach the standard, the current dynamic event e is stored in the global dynamic event group E, and a new dynamic event e is read to execute step S200.

[0080] Specifically, in the example, considering that the vehicle in the scene may have a certain deflection angle, if the density of the local dynamic event group F does not reach the threshold, the dodecahedron spatio-temporal neighborhood can be rotated clockwise in the x-y plane to adjust the spatial position, for example: rotating 15° each time, and again determining whether the density threshold is reached.

[0081] If the dodecahedron is rotated 12 times and the density of the extracted local dynamic event group F still does not reach the threshold N, the current dynamic event e is stored in the global dynamic event group E, and the next dynamic event is read from the bionic dynamic event stream, and the step S200 is repeated.

[0082] Furthermore, in order to prevent repeated positioning, in step S200, the step further comprises: step S230 deleting all dynamic events in the local dynamic event group F judged as the candidate target region from the global dynamic event group E to update the global dynamic event group E, thereby preventing repeated positioning.

[0083] Step S300 maps the candidate target region into a binary image G; according to the time and position information of the dynamic events in the candidate target region, a new label image is cut out from the corresponding scene label image.

[0084] In step S300, the step of mapping the candidate target region into a binary image G comprises:

[0085] Step S310 initializes a binary image P with resolution of L L, and sets the pixel value of each position to 0, wherein L is the length of the square candidate target region.

[0086] Step S320 reads all dynamic events in the local dynamic event group F of the candidate target region, and maps them into the binary image P according to the position information of each dynamic event, wherein the mapping function is:

[0087]

[0088] wherein is the position coordinate of each dynamic event in the candidate target region, is the coordinate of the dynamic event e at the center point, is the time information of the dynamic event at the center point, is the new coordinate of each dynamic event after mapping.

[0089] Step S330 sets the pixel value of all dynamic events mapped in P to 255, and thus obtains the binary image G as shown in Figure 7 .

[0090] Furthermore, the example steps of cutting out the new label image comprise:

[0091] Step S340 positions the center event of the candidate target region to the scene label image with the closest time stamp according to the time information of the center dynamic event e of the candidate target region.

[0092] Step S350 cuts out the corresponding candidate target region from the scene label image according to the length L of the square and the position information of the center dynamic event e of the candidate target region in the spatial neighborhood of the candidate target region, and thus obtains a new label image with size of L L.

[0093] Wherein the new label map corresponds to the candidate target region one by one, if the new label map contains a vehicle, it represents that the corresponding candidate target region contains a vehicle, and the region is a positive sample, otherwise it is a negative sample. The new label map is as shown in Figure 8 .

[0094] Step S400 establishes a target contour recognition network to extract the moving target contour in the new label map, as shown in Figure 9 . On the other hand, a target detection network also needs to be established to input the binary image G and the moving target contour information, and to perform fusion training to determine whether it is a moving target such as a vehicle in the binary image G. Wherein the vehicle contour information in the training label map is fused in order to make the network learn more target features.

[0095] Step S500 establishes a global target recognition network to take all dynamic events in the candidate target region as a label and take the global dynamic event group E as an input to complete training to output moving target information.

[0096] Specifically, the global target recognition network is mainly used to detect vehicle targets from the 3D point cloud-shaped global dynamic event group E. For this purpose, in the training process, the candidate target region in the aforementioned positive sample contains vehicle dynamic events, which can be used as a label to train the network, thereby solving the problem puzzled by the prior art.

[0097] Finally, the global dynamic event group E obtained by the foregoing is used as an input to utilize the trained global target recognition network to detect vehicle targets from the global dynamic event group E to obtain the center point dynamic event information of the vehicle target.

[0098] Therefore, through the scheme of the foregoing example, on the one hand, the bionic dynamic signal can be converted into a candidate target region for the target contour recognition and target detection neural network to detect whether the region is a vehicle, and on the other hand, through the global target recognition network, vehicle information (i.e. position coordinate information of the vehicle in the x-y plane) can be directly detected from the global dynamic event group E, thereby realizing double detection of the aerial moving target.

[0099] Corresponding to the foregoing method, as shown in Figure 10 , the present application also provides a double detection system of the aerial moving target, which comprises:

[0100] A storage unit for storing the program comprising the steps of the double detection method of the aerial moving target according to any one of the foregoing embodiments, for the positioning unit, the first detection unit, the training set processing unit, the second detection unit, and the like to be executed in time.

[0101] A positioning unit for buffering The time period is a dynamic event stream of bionics, a global dynamic event group E is established, and is updated in a loop; a polygon modeling of a moving target is constructed as a spatial neighborhood, A spatiotemporal neighborhood is constructed for the length, and a local dynamic event group F in E is extracted; the density of the local dynamic event group F is judged, when the density reaches a standard, the spatiotemporal neighborhood is deformed to a cube as a candidate target region, and the local dynamic event group F in the candidate target region is deleted from the global dynamic event group E.

[0102] A training set processing unit is configured to map the candidate target region into a binary image G according to coordinate information; find a scene label image corresponding to a time stamp according to time information of e in the candidate target region, and mark the corresponding candidate target region in the scene label image according to the position information of e, and cut out a new label image.

[0103] A first detection unit is configured to identify whether the binary image G is a moving target according to the new label image and the binary image G.

[0104] A second detection unit is configured to take all dynamic events in the candidate target region corresponding to the positive sample as a label, and take the global dynamic event group E as an input to complete training of a global target recognition network, and then identify moving target information in the dynamic event.

[0105] In the preferred example, the detection unit includes: a target contour recognition network configured to extract a moving target contour in the new label image; and a target detection network configured to take the binary image G and the moving target contour information as an input to perform fusion training to determine whether the binary image G is a moving target.

[0106] Corresponding to the above method, the application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the aerial moving target double detection method according to any one of the above embodiments.

[0107] In summary, the aerial moving target double detection method and system, and the storage medium provided by the application can process bionic dynamic events one by one, fully utilize the asynchronous sparsity of bionic dynamic events, and reduce the redundancy of the overall algorithm; at the same time, the polyhedron modeling scheme is innovatively designed to locate a candidate target from a bionic dynamic event stream, thereby quickly eliminating useless information in the scene and focusing on the region with moving targets, thereby improving the response speed of the algorithm.

[0108] In addition, the positioning result of the candidate target can solve the problem of lack of labels when bionic dynamic events are used for neural network training, so that bionic dynamic events can be used for neural network training and detection, thereby fully utilizing the advantages of bionic dynamic events in the field of motion detection and improving the real-time performance and detection accuracy of the overall algorithm.

[0109] The preferred embodiments of the application disclosed above are only used to illustrate the present application. The preferred embodiments are not meant to limit the present application to any particular detailed embodiments. Obviously, many modifications and variations are possible in light of the above teachings. It is intended that the scope of the application be limited only by the widest range of equivalency to the claims below and their equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of the present application.

[0110] Those skilled in the art can understand that, in addition to implementing the system, device, unit and each module thereof provided by the present application in the form of pure computer readable program code, the same program can also be realized by logically programming the method steps to the form of logic gate, switch, application specific integrated circuit, programmable logic controller and embedded microcontroller, etc. Therefore, the system, device and each module thereof provided by the present application can be considered as a hardware component, and the modules included therein for realizing various programs can also be considered as structures in the hardware component; the modules for realizing various functions can also be considered as both software programs for realizing methods and structures in the hardware component.

[0111] In addition, all or part of the steps of the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a program stored in a storage medium, including a plurality of instructions for causing a single-chip microcomputer, chip or processor to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0112] In addition, various different embodiments of the embodiments of the present application can also be combined arbitrarily, as long as they do not deviate from the idea of the embodiments of the present application, and they should also be considered as disclosed by the embodiments of the present application.

Claims

1. A method for detecting aerial moving targets, comprising the steps of: S100, obtaining a bionic dynamic event stream, establishing a cyclically updated global dynamic event group E, polygonally modeling moving targets in a scene label graph, and constructing a spatio-temporal neighborhood to extract a local dynamic event group F in E, comprising the steps of: Step S110 buffering Bionic dynamic event flow in time period, establish global dynamic event group E; Step S120 reads a new dynamic event e from the bionic dynamic event stream, finds the minimum value of the time of all dynamic events in the global dynamic event group E If the time information t of e is used as a reference, and if t- Then the dynamic events in the global dynamic event group E whose absolute value of the difference between the time stamp and t is greater than need to be deleted to refresh the global dynamic event group E;​ S130, selecting a single moving target from the scene label graph, and constructing an inscribed circle with the length of the moving target as the diameter and the center coordinates as the center; Step S140 constructs a polygon in the circumcircle to surround the moving target; the distance and angle of each vertex of the polygon relative to the center coordinates of the moving target are calculated to construct a corresponding relationship function , modeling is completed; Step S150 centers on the location information of the dynamic event e, and constructs a polygon centered on e as a spatial neighborhood according to a relationship function . Step S160 selects in time dimension For length, stretch the polygonal spatial neighborhood to construct a spatio-temporal neighborhood; S200, when the density of the local dynamic event group F meets the standard, inflating the spatio-temporal neighborhood to a cube to serve as a candidate target region; S300, mapping the candidate target region into a binary image G, and extracting a new label graph in the corresponding scene label graph according to the time and position information of the dynamic events in the candidate target region; S400, establishing a target contour recognition network to extract the contour of the moving target in the new label graph, and establishing a target detection network to take the binary image G and the contour information of the moving target as inputs, and perform fusion training to determine whether the binary image G is a moving target; S500, establishing a global target recognition network to take all dynamic events in the candidate target region as labels and the global dynamic event group E as inputs, and complete training to output moving target information. 2.The method of claim 1, wherein in step S200, the steps further comprise: S210, when the density of the local dynamic event group F does not meet the standard, adjusting the spatial position of the spatio-temporal neighborhood within a preset number of times until the standard is met, and then executing step S300; S220, if the standard is still not met, storing the current dynamic event e in the global dynamic event group E, and reading a new dynamic event e to execute step S200. 3.The method of claim 1, wherein in step S200, the steps further comprise: S230, deleting all dynamic events in the local dynamic event group F that is determined to be the candidate target region from the global dynamic event group E to update the global dynamic event group E. 4.The method of claim 1, wherein in step S300, the step of mapping the candidate target region into a binary image G comprises: Step S310 initializes a binary picture P of resolution L L, where L is the length of the square candidate target region, and each pixel value is 0. S320, reading all dynamic events in the local dynamic event group F of the candidate target region, and mapping them to a binary image P according to the position information of each dynamic event; S330, assigning the position pixel values of all dynamic events in P to 255 to obtain the binary image G. 5.A system for detecting aerial moving targets, comprising: a storage unit for storing programs comprising the steps of the method for detecting aerial moving targets according to any one of claims 1 to 4, for the positioning unit, the first detection unit, the training set processing unit, the second detection unit, and the timely calling execution; Positioning unit for caching Bionic dynamic event flow in time period, establish global dynamic event group E and update cyclically; construct spatial neighborhood with new read dynamic event e as center and polygon modeling of moving target, Space-time neighborhood for length group, extract local dynamic event group F in E; judge density of local dynamic event group F, when up to standard, expand space-time neighborhood to cube as candidate target region, and delete local dynamic event group F in this candidate target region from global dynamic event group E; a training set processing unit for mapping the candidate target region into a binary image G according to the coordinate information; According to the time information of e in the candidate target region, a scene label graph corresponding to the timestamp is found, and according to the position information of e, a corresponding candidate target region is marked in the scene label graph, and a new label graph is intercepted; The first detection unit is configured to identify whether the binary image G is a moving target according to the new label graph and the binary image G; The second detection unit is configured to take all dynamic events in the candidate target region corresponding to the positive sample as a label, take the global dynamic event group E as an input, and complete training of a global target recognition network to identify moving target information in the dynamic events.

6. The dual detection system for aerial moving targets according to claim 5, wherein the detection unit comprises: a target contour recognition network configured to extract a moving target contour in the new label graph; a target detection network configured to take the binary image G and the moving target contour information as an input, perform fusion training, and determine whether the binary image G is a moving target.

7. A computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the dual detection method for aerial moving targets according to any one of claims 1 to 4.

Images