Target information continuous generation method and system

Abstract

The invention provides a method and a system for continuously generating target information. The method comprises the following steps: constructing a detection model for detecting a to-be-detected target; obtaining a sequence remote sensing image; determining a first image in which a suspected to-be-detected target appears for the first time from the sequence remote sensing image, and detecting a first target in the first image through a detection model; extracting a plurality of continuous second images in which suspected to-be-detected targets appear after the first image from the sequence remote sensing image, and detecting a plurality of second targets in the plurality of second images through a detection model to form a second target set; calculating the suspected degree of each second target in the second target set and the first target, and screening out a part of second targets from the plurality of second targets according to the suspected degree to form a high-suspected-degree target chain; performing neighborhood constraint processing on a second image corresponding to each second target in the high-suspected-degree target chain to obtain a plurality of third images; and counting the number of times that the second target appears in the plurality of third images, and judging whether the first target is the to-be-detected target or not.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and more particularly, to a method and system for continuously generating target information. Background Art

[0002] With the development of computer technology and remote sensing technology, the spatial resolution and temporal resolution of remote sensing images have been greatly improved, providing richer information for image analysis using remote sensing images and time series. The target information continuous generation technology based on sequence remote sensing images has a wide range of applications in fields such as traffic management.

[0003] Target detection based on image sequences is a key area of ​​future development in ocean surveillance, and this is where the use of video and visible light remote sensing satellites for target observation excels. In the civilian sector, remote sensing technology and computer vision techniques can be used to analyze features such as shape, texture, and color in remote sensing images, distinguishing between background and target, and extracting useful target information. This serves applications such as traffic detection, environmental monitoring, ocean inspection, terrain mapping, and resource exploration.

[0004] There is also target detection through single-frame image and trajectory extraction technology. These two technologies have fast detection speeds, but do not fully utilize the multi-frame time series information of the image, and have high false alarm and missed alarm rates. There is still a lot of room for improvement in practical applications.

[0005] Moreover, optical remote sensing images are easily affected by clouds, fog, and lighting, and there are very few pixels of targets such as ships in the images, which are weak targets. Therefore, the difficulty of target detection and tracking in optical remote sensing sequence images lies mainly in the fact that in complex backgrounds such as dense cirrus clouds, when the target movement speed changes or the wind speed and direction change, the characteristics of weak targets change accordingly, which makes it easy to have problems such as missed tracking and wrong tracking.

[0006] Existing orbiting satellites can capture high-temporal-resolution remote sensing image sequences and even videos. For example, the Gaofen-4 satellite's panchromatic optical channel can capture video-like image sequences with a frame rate of as fast as 5 seconds. Currently, multi-frame correlation technology is mostly used in natural image target tracking algorithms, but research on using multi-frame correlation for ship target detection in remote sensing images is still limited. Therefore, it is feasible to use continuous multi-frame images for target detection to improve performance. Summary of the Invention

[0007] In view of the above problems, the present invention provides a method and system for continuously generating target information.

[0008] On the one hand, the present invention provides a method for continuously generating target information, including: constructing a detection model for detecting a target to be detected; obtaining a sequence of remote sensing images to be processed, the sequence of remote sensing images including a continuous plurality of frames; determining a first image in which a suspected target to be detected first appears from the sequence of remote sensing images, and detecting the first target in the first image through the detection model; extracting a plurality of consecutive second images in which the suspected target to be detected appears after the first image from the sequence of remote sensing images, and detecting a plurality of second targets in the plurality of second images through the detection model to form a second target set; calculating the degree of suspicion between each second target in the second target set and the first target, and screening out some second targets from the plurality of second targets according to the degree of suspicion to form a high-suspicion target chain; performing neighborhood constraint processing on the second image corresponding to each second target in the high-suspicion target chain to obtain a plurality of third images; and counting the number of times the second target appears in the plurality of third images to determine whether the first target is the target to be detected.

[0009] According to an embodiment of the present invention, a YOLO model is selected and optimized using the SENet channel attention mechanism. A BiFPN module is added to the optimized YOLO model, and the convolutional layer in the added YOLO model is replaced with a fractal convolution module to form a detection model.

[0010] According to an embodiment of the present invention, a temporal correlation measurement value of the target to be detected is obtained from the detection model; and features of the plurality of second targets in the plurality of second images are extracted using the temporal correlation measurement value.

[0011] According to an embodiment of the present invention, a sequence signal is sent, and the sequence signal sent by the second target in the previous frame is associated with the sequence signal sent by the second target in the next frame; for multiple second targets, the sequence signal of the previous second target is associated with the sequence signal of the next second target to form an associated sequence; and the multiple associated sequences are determined as a second target set.

[0012] According to an embodiment of the present invention, a quantitative characterization model is established based on the distribution law of the characteristics of the second target; and the degree of suspicion between each associated sequence in the set and the time series correlation metric value is determined based on the quantitative characterization model.

[0013] According to an embodiment of the present invention, when the sequence signals of any two second targets among multiple second targets are the same, the sequence signals of the any two second targets are associated to form an associated sequence; the associated sequence signals are identified as sub-chains under the same second target parent chain; and the multiple sub-chains are merged and judged to obtain sub-chains with different suspicion levels, forming a high-suspicion target chain.

[0014] According to an embodiment of the present invention, when multiple sub-chains meet specified conditions, the sub-chain with the highest suspicion is retained; when multiple sub-chains do not meet the specified conditions, the sub-chain is separated and becomes a new second target parent chain.

[0015] According to an embodiment of the present invention, data processing is performed on a high-suspect target chain to suppress the complex background of the high-suspect target chain; filtering is performed on the processed high-suspect target chain to remove clutter from the complex background; neighborhood convolution is performed on the high-suspect target chain after clutter removal to obtain information about a second target; calculations are performed on each of two adjacent second images using the second target information to obtain the tracks of multiple second targets; and constraint processing is performed on the tracks of the multiple second targets to obtain multiple third images.

[0016] According to an embodiment of the present invention, a plurality of third images are searched to detect whether a second target exists in each third image, wherein the second target exists in the third image when the track of the second target satisfies a constraint condition; the number of times the second target appears in the plurality of third images is counted and compared with a set threshold; when the number is greater than or equal to the threshold, the first target is the target to be measured.

[0017] Another aspect of the present invention provides a target information continuous generation system, including: a target detection module, used to collect features of the target to be measured and construct a detection model for detecting the target to be measured; a target extraction module, used to obtain a sequence of remote sensing images to be processed, the sequence of remote sensing images including multiple consecutive frames of images, determine the first image in which the suspected target to be measured first appears from the sequence of remote sensing images, and extract multiple second images consecutive after the first image from the sequence of remote sensing images; a target processing module, used to perform suspicion degree calculation and neighborhood constraint processing on the multiple second images to obtain multiple third images; a target judgment module, used to perform threshold comparison on the multiple third images to confirm whether the first target in the first image that appears first as the suspected target to be measured is the target to be measured.

[0018] The target information continuous generation method and system provided by the present invention can achieve the following beneficial effects:

[0019] Taking into account the high resolution characteristics of visible light images, a fast infrared small target detection algorithm based on the background difference method is designed for the pure sea or pure cloud background in infrared images. Combined with the motion characteristics of aircraft targets, an inter-frame target association algorithm based on multi-frame target mapping is adopted to extract the target track and perform false alarm suppression. This solves the problems of target loss, missed tracking, and wrong tracking due to invalid feature matching, effectively reduces the false alarm rate and missed alarm rate, and improves the tracking speed and robustness in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0021] Figure 1 A flowchart of a method for continuously generating target information according to an embodiment of the present invention is schematically shown;

[0022] Figure 2 Schematically shows a principle diagram of target information image processing according to an embodiment of the present invention;

[0023] Figure 3 The flowchart of constructing a detection model for detecting an object to be detected according to an embodiment of the present invention is schematically shown;

[0024] Figure 4 The figure schematically shows a structural diagram of a Dense-Yolo network model structure according to an embodiment of the present invention.

[0025] Figure 5 The schematic diagram shows the principle of fractal convolution module and conventional convolution layer feature collection according to an embodiment of the present invention.

[0026] Figure 6 A schematic diagram illustrating a principle of forming a target association sequence according to an embodiment of the present invention is shown;

[0027] Figure 7 The following schematically shows a principle diagram of merging judgment of subchains in a target chain according to an embodiment of the present invention;

[0028] Figure 8 A flowchart of processing a second image to obtain a third image according to an embodiment of the present invention is schematically shown;

[0029] Figure 9 The block diagram of the target information continuous generation system according to an embodiment of the present invention is schematically shown.

[0030] Description of reference numerals:

[0031] 100 - first image; 200 - second image; 300 - third image; 400 - sequence signal; 500 - computer; 600 - mother chain; 601 - daughter chain No. 1; 602 - daughter chain No. 2; 603 - daughter chain No. 3; 604 - daughter chain No. 4. DETAILED DESCRIPTION

[0032] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0033] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise," "include," etc. used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0034] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0035] Before describing specific embodiments of the present invention in detail, technical terms are first explained to facilitate a better understanding of the present invention.

[0036] Sequential remote sensing images: refers to multiple frames of images acquired by satellites, drones or other remote sensing platforms at consecutive time points. These images are correlated in time, space and spectral dimensions and can be used for tasks such as dynamic monitoring, target tracking and change detection.

[0037] Neighborhood constraint: refers to the use of the correlation of the local area (neighborhood) around the target object in data processing or model optimization to impose restrictions or guiding conditions on the results to improve accuracy, continuity or physical rationality, so that the attributes or behaviors of the target object should maintain a certain correlation with its neighboring objects.

[0038] In view of this, the present invention provides a method and system for continuously generating target information.

[0039] Figure 1 A flowchart of a method for continuously generating target information according to an embodiment of the present invention is schematically shown; Figure 2 The principle diagram of target information image processing according to an embodiment of the present invention is schematically shown.

[0040] like Figure 1 and Figure 2 As shown, the method for continuously generating target information according to this embodiment includes steps S1 to S7.

[0041] In step S1, a detection model for detecting an object to be detected is constructed.

[0042] For example, the target to be measured can be an aircraft. In view of the fact that different types of aircraft have different sizes, a detection model based on deep learning of optical remote sensing images is constructed by combining local environmental information perception, difficult sample mining, BiFPN multi-scale fusion strategy, and smaller target protection mechanisms such as deconvolution, dense connection, and multi-scale receptive field mechanism.

[0043] For example, the detection model may adopt the YOLO model, Faster R-CNN, etc.

[0044] In step S2, a sequence of remote sensing images to be processed is obtained, where the sequence of remote sensing images includes multiple continuous frames of images.

[0045] In step S3, a first image 100 in which the suspected target to be detected first appears is determined from the sequence of remote sensing images, and the first target in the first image 100 is detected using a detection model.

[0046] The first image 100 where the suspected target to be detected first appears is located from the sequence remote sensing image. The first image 100 is quickly screened through a temporal difference or anomaly detection algorithm, and the detection model is used to analyze the first image 100 to identify and locate the first target information therein. The information includes the output target category, location, and confidence level.

[0047] In step S4, a plurality of consecutive second images 200 in which suspected targets appear after the first image 100 are extracted from the sequence of remote sensing images, and a plurality of second targets in the plurality of second images 200 are detected by the detection model to form a second target set.

[0048] Extract multiple consecutive frames of images after the first appearance of the target from the sequence remote sensing image to ensure temporal continuity and spatial coverage consistency, analyze the second image 200 frame by frame, detect multiple second targets therein, output the target position, category and confidence level, and form a second target set.

[0049] In step S5, the suspicion degree between each second target in the second target set and the first target is calculated, and some second targets are selected from the plurality of second targets according to the suspicion degree to form a high suspicion target chain.

[0050] Extract the features of each target in the first target and the second target set, such as appearance, texture, spectral characteristics, motion vector, etc., calculate the suspicion degree through similarity measurement, combine the target motion model such as Kalman filter predicted position, or temporal context such as optical flow trajectory consistency, adjust the suspicion degree weight, retain the second target with a suspicion degree higher than the set value, and eliminate high-suspected targets that overlap in space or time to avoid redundancy.

[0051] For example, the similarity metric may be one or more of cosine similarity, IoU, and HOG feature matching.

[0052] In step S6 , neighborhood constraint processing is performed on the second image 200 corresponding to each second target in the high-suspected target chain to obtain a plurality of third images 300 .

[0053] With each highly suspected target as the center, a surrounding area of ​​a fixed range or an adaptive range (based on the target size) is extracted. The target positions of the previous and next frames are combined, the neighborhood boundaries are dynamically adjusted, the pixel labels within the neighborhood are optimized, a smooth transition between the target and the background is ensured, the target features within the neighborhood are enhanced, guided filtering or wavelet transform is used to suppress noise, the area within the neighborhood is processed, the other parts of the image are kept unchanged, and multiple third images 300 are output to enhance the saliency of the target in a complex background.

[0054] In step S7 , the number of times the second target appears in the plurality of third images 300 is counted to determine whether the first target is the target to be detected.

[0055] During the detection process, the image in which the suspected target first appears is used as the initial image, and the position of the first target is used as the initial position. The next N detection result images are searched, and each image is checked for the presence of a second target. If the cumulative number of times the second target is detected in the image is greater than or equal to the set threshold d, the first target in the initial image is considered to be the true target. If it is below the threshold d, the first target is excluded.

[0056] Figure 3 A flowchart of constructing a detection model according to an embodiment of the present invention is schematically shown; Figure 4 A structural diagram schematically shows a Dense-Yolo network model structure according to an embodiment of the present invention; Figure 5 The schematic diagram shows the principle of fractal convolution module and conventional convolution layer feature collection according to an embodiment of the present invention.

[0057] like Figure 3 、 Figure 4 and Figure 5 As shown, the construction of a detection model for detecting a target to be detected according to this embodiment includes steps S11 to S12.

[0058] In step S11, the YOLO model is selected and the SENet channel attention mechanism is used to optimize the YOLO model.

[0059] In view of the characteristics of different sizes of different types of aircraft and the application requirements of rapid target detection, a deep learning target detection algorithm with better scale adaptability, Dense-Yolo, is proposed based on the highly time-effective YOLO network and targeted improvements. The Dense-Yolo network model structure adopts the DenseNet network. Based on the idea of ​​dense network connection of the DenseNet network, the residual network structure Resunit in the two CSP1_3 modules in the backbone network is replaced with a custom Denseblock module. The squeeze and excitation method in SENet is used to optimize the network structure, thereby explicitly modeling the interdependence between feature channels.

[0060] For example, the difference between the DenseNet network and the residual network is that the Nth layer output of the Resunit module comes from the The output of the layer plus the The nonlinear transformation of the output of the layer is expressed as follows:

[0061]

[0062] Where, represents the Nth layer output, represents the nonlinear transformation of the Nth layer, Indicates the Layer output.

[0063] For example, the specific formula of the Denseblock module is expressed as:

[0064]

[0065] Where, represents the Nth layer output, represents the nonlinear transformation of the Nth layer, Before The output feature map of the layer is fused with channels.

[0066] In step S12, the BiFPN module is added to the optimized YOLO model, and the convolutional layer in the added YOLO model is replaced with a fractal convolution module to form a detection model.

[0067] The BiFPN module has been added to improve the scale compatibility of small-scale objects. FPN is an enhancement to the traditional CNN network's representation and output of image information. Its purpose is to improve the CNN network's feature extraction method, so that the final output features better represent information from all dimensions of the input image. BiFPN builds on the complex bidirectional fusion of the FPN network by adding a weighting mechanism, assigning weights to different scales. Traditionally, features of different scales are simply stacked, while BiFPN allows the network to independently learn the weights of different input features, more efficiently fusing features of different scales in both directions.

[0068] The fractal convolution module improves the prediction and regression of aircraft targets of various models with varying aspect ratios. When using conventional convolutional layers in the YOLO model for feature collection, the features collected on narrow edges are relatively sparse. This results in insufficient feature information for aircraft model prediction and bounding box regression on narrow edges, making it difficult to support good prediction and regression. Unlike conventional convolutional layers, the fractal convolution module maintains a consistent feature sampling rate for aircraft features in both width and height dimensions, ensuring the proper collection of key feature information from aircraft targets.

[0069] For example, the regression layer in the fractal convolution module accelerates the convergence speed of the network to a certain extent. The convergence speed formula is expressed as:

[0070]

[0071]

[0072] Where, represents the convergence speed in the width direction, represents the convergence speed in the high direction, represents the convolution step size on the i-th feature layer, Represents the predicted value on the i-th feature layer in the width direction, Represents the predicted value on the i-th feature layer in the high direction, Indicates the actual value in the width direction, represents the predicted value in the width direction, Indicates the actual value of the high direction, Indicates a predicted value in the high direction.

[0073] Figure 6 The principle diagram of forming a target association sequence according to an embodiment of the present invention is schematically shown.

[0074] like Figure 6As shown, a sequence signal 400 is sent by a computer 500, and each second target sends a unique sequence signal 400 in consecutive frames. The sequence signal 400 includes metadata such as target name, position, speed, and confidence. The sequence signal 400 sent by the second target in the previous frame is associated with the sequence signal 400 sent by the second target in the next frame. For multiple second targets, the sequence signal 400 of the previous second target is associated with the sequence signal 400 of the next second target. For each second target, its cross-frame sequence signals 400 are spliced ​​in chronological order to form multiple associated sequences, which constitute a second target set.

[0075] For example, the sequence initiated by the previous sequence can filter and associate the second target of the subsequent frame based on the characteristics of the limited speed, energy, speed direction and stability of size changes. It can also associate the second target after a certain number of frames, and update the data of the sequence with association points.

[0076] Figure 7 The diagram schematically shows the principle of merging sub-chains in a target chain according to an embodiment of the present invention.

[0077] like Figure 7 As shown, when the sequence signals 400 of any two of the multiple second targets are identical, the sequence signals 400 of the two second targets are correlated to form a correlated sequence, and the correlated sequence signals 400 are identified as subchains under the same second target parent chain 600. Multiple subchains (subchain 1 601, subchain 2 602, subchain 3 603, and subchain 4 604) are merged and judged to obtain subchains with different suspicions, forming a high-suspicion target chain. If the multiple subchains (subchain 1 601, subchain 2 602, subchain 3 603, and subchain 4 604) meet specified conditions, the subchain with the highest suspicion is retained. If the multiple subchains do not meet the specified conditions, the subchains are separated and become new second target parent chains 600.

[0078] For example, the sequence is judged by the discontinuity threshold and the "M select N" condition. If any of the discarding conditions corresponding to the above judgments are met, the sequence is discarded. A quantitative representation model is established by the feature distribution rules of different dimensions of the second target (displacement, speed, energy, distance between measured position and predicted position, etc.) to measure the similarity between the features of the suspected target association sequence and the quantitative values ​​of the second target features.

[0079] For example, in the case where the same sequence is associated with multiple second targets, the newly associated sequences are identified as different sub-chains under the same parent chain 600 (sub-chain No. 1 601, sub-chain No. 2 602, sub-chain No. 3 603, and sub-chain No. 4 604). The sub-chains under the same parent chain 600 that subsequently re-intersect at the same point are merged and judged. If the conditions are met, only the sub-chain with the highest suspicion (sub-chain No. 3 603) is retained, and the remaining sub-chains (sub-chain No. 1 601, sub-chain No. 2 602, and sub-chain No. 4 604) are eliminated. If the conditions are not met and the generated sub-chain frame sequence number is greater than the tolerance threshold compared to the current frame sequence number, the sub-chains (sub-chain No. 1 601, sub-chain No. 2 602, sub-chain No. 3 603, and sub-chain No. 4 604) are separated to become new independent parent chains 600. There are three forms:

[0080] (a) Subchain No. 1 601 serves as the new parent chain 600, and subchain No. 2 602, subchain No. 3 603, and subchain No. 4 604 are all subchains under this new parent chain 600;

[0081] (b) Subchain No. 1 601 serves as the new mother chain 600, subchain No. 2 602 serves as the subchain of subchain No. 1 601, subchain No. 3 603 serves as the new mother chain 600, and subchain No. 4 604 serves as the subchain of subchain No. 3 603;

[0082] (c) Sub-chain No. 1 601 , sub-chain No. 2 602 , sub-chain No. 3 603 , and sub-chain No. 4 604 all serve as new mother chains 600 .

[0083] Figure 8 The flowchart of processing the second image 200 to obtain the third image 300 according to an embodiment of the present invention is schematically shown.

[0084] like Figure 8 As shown, according to an embodiment of the present invention, processing the second image 200 to obtain the third image 300 includes steps S61 to S64.

[0085] In step S61, data processing is performed on the highly suspected target chain to suppress the complex background of the highly suspected target chain.

[0086] In step S62, the processed high-suspect target chain is filtered to remove clutter of the complex background.

[0087] By performing minimum relative grayscale gradient filtering on the target bidirectional oversampled image, noise can be effectively reduced. The purpose of minimum relative grayscale gradient filtering is to filter out high-intensity noise and drastically changing background clutter.

[0088] For example, filtering can also be done by: (1) guided filtering, which uses a guiding image (such as the target neighborhood) to constrain the filtering process and preserve edges; (2) bilateral filtering, which combines spatial distance and pixel value similarity for weighted smoothing; and (3) non-local mean filtering, which uses globally similar blocks of the image for weighted averaging.

[0089] In step S63, neighborhood convolution processing is performed on the high-suspected target chain after clutter removal to obtain information about the second target.

[0090] While performing noise reduction processing on the second image 200, neighborhood convolution processing is performed on the second target, and the target pixel value is replaced by the convolution result. The more bright pixels there are around the second target pixel, the higher its brightness gain after convolution. The convolution window is determined so that the bright pixels closer to the center of the window contribute greater gain, which helps to retain the second target information.

[0091] In step S64 , each two adjacent frames of the second image 200 are calculated using the information of the second target to obtain the tracks of multiple second targets, and the tracks of the multiple second targets are constrained to obtain multiple third images 300 .

[0092] In the infrared image sequence, the arrangement interval is designed so that the target point spacing formed by the second target in the air on the difference image between two adjacent frames of the sequence and the background meets the neighborhood constraint condition. The distance between the same target in adjacent frames is related to the speed range of the target.

[0093] For example, target similarity (r) is expressed as the correlation coefficient of the small target area detected in two sequence images, and its formula is expressed as

[0094]

[0095] Where, represents the mean value of region A, represents the mean of region B, Indicates that the coordinates in area A are ( , ), Indicates that the coordinates in area B are ( , ) of the observations.

[0096] In order to avoid the influence of the target on the calculation results of the mean and mean square error, the central area of ​​the neighborhood image is first removed, and the mean and variance calculations are only performed on the area outside the removed target.

[0097] Based on the method disclosed in the above embodiment, the present invention also provides a target information continuous generation system. Figure 9 The system is described in detail.

[0098] Figure 9 The block diagram of the target information continuous generation system according to an embodiment of the present invention is schematically shown.

[0099] like Figure 9 As shown, the target information continuous generation system 700 according to this embodiment includes a target detection module 710 , a target extraction module 720 , a target processing module 730 and a target judgment module 740 .

[0100] The target detection module 710 is used to collect features of the target to be detected and build a detection model for detecting the target to be detected.

[0101] The target extraction module 720 obtains a sequence of remote sensing images to be processed, the sequence of remote sensing images including multiple consecutive frames of images, determines the first image 100 in which the suspected target to be detected first appears from the sequence of remote sensing images, and extracts multiple second images 200 consecutive after the first image 100 from the sequence of remote sensing images.

[0102] The target processing module 730 is configured to perform suspicion calculation and neighborhood constraint processing on the plurality of second images 200 to obtain a plurality of third images 300 .

[0103] The target determination module 740 performs a threshold comparison on the plurality of third images 300 to determine whether the first target that appears first in the first image 100 as a suspected target to be detected is the target to be detected.

[0104] It should be noted that the implementation method of the system part is similar to the implementation method of the above-mentioned method part, and the technical effects achieved are also similar. For specific details, please refer to the above-mentioned method implementation method part, which will not be repeated here.

[0105] According to embodiments of the present invention, any multiple of the target detection module 710, target extraction module 720, target processing module 730, and target determination module 740 may be combined into a single module, or any one of these modules may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in a single module. According to embodiments of the present invention, at least one of the target detection module 710, target extraction module 720, target processing module 730, and target determination module 740 may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or may be implemented in hardware or firmware through any other reasonable means of circuit integration or packaging, or may be implemented in any one of software, hardware, and firmware, or any appropriate combination of these. Alternatively, at least one of the target detection module 710, the target extraction module 720, the target processing module 730, and the target judgment module 740 can be at least partially implemented as a computer program module, and when the computer program module is executed, the corresponding function can be performed.

[0106] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architectures, functions and operations of the devices and methods according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0107] Those skilled in the art will appreciate that the features described in the various embodiments of the present invention may be combined or coupled in various ways, even if such combinations or couplings are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention may be combined or coupled in various ways without departing from the spirit and teachings of the present invention. All such combinations or couplings fall within the scope of the present invention.

[0108] The above describes embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be advantageously used in combination. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A method for continuously generating target information, characterized in that: include: Constructing a detection model for detecting the target; Acquiring a sequence of remote sensing images to be processed, wherein the sequence of remote sensing images includes a continuous plurality of frames of images; Determining a first image in which the suspected target to be detected first appears from the sequence of remote sensing images, and detecting a first target in the first image using the detection model; Extracting a plurality of consecutive second images in which the suspected target to be detected appears after the first image from the sequence of remote sensing images, and detecting a plurality of second targets in the plurality of second images using the detection model to form a second target set; calculating a suspicion degree between each second target in the second target set and the first target, and screening out some second targets from the plurality of second targets based on the suspicion degrees to form a high-suspicion target chain; performing neighborhood constraint processing on the second image corresponding to each second target in the high-suspected target chain to obtain a plurality of third images; The number of times the second target appears in the plurality of third images is counted to determine whether the first target is the target to be detected.

2. The method according to claim 1, characterized in that Building a detection model for detecting the target includes: Select the YOLO model and optimize it using the SENet channel attention mechanism; The BiFPN module is added to the optimized YOLO model, and the convolutional layer in the added YOLO model is replaced with a fractal convolution module to form the detection model.

3. The method according to claim 1, characterized in that Detecting the plurality of second targets in the plurality of second images by using the detection model comprises: Obtaining a temporal correlation metric value of the target to be tested from the detection model; The features of the plurality of second objects in the plurality of second images are extracted using the temporal correlation metric value.

4. The method according to claim 3, characterized in that The characteristic of the second target includes a sequence signal, and forming the second target set further includes: sending the sequence signal, wherein the sequence signal sent by the second target in the previous frame is associated with the sequence signal sent by the second target in the next frame; For the plurality of second targets, correlating the sequence signal of a preceding second target with the sequence signal of a succeeding second target to form a correlation sequence; A plurality of the associated sequences are determined as the second target set.

5. The method according to claim 4, characterized in that Calculating the suspicion between each second target in the second target set and the first target includes: Establishing a quantitative characterization model based on the distribution law of the characteristics of the second target; The degree of suspicion between each associated sequence in the set and the temporal correlation metric value is determined by using the quantitative representation model.

6. The method according to claim 4, characterized in that The method further comprises screening out some second targets from the plurality of second targets according to the suspicion levels to form a high-suspect target chain, wherein the target chain includes a plurality of parent chains and at least one child chain subordinate to each parent chain, and further comprises: When the sequence signals of any two second targets among the plurality of second targets are the same, correlating the sequence signals of the any two second targets to form a correlated sequence; Identifying the associated sequence signals as daughter chains under the same second target parent chain; The plurality of sub-chains are combined and judged to obtain sub-chains with different suspicion levels, thereby forming a target chain with a high suspicion level.

7. The method according to claim 6, characterized in that The merging judgment of the sub-chains includes: When multiple subchains meet the specified conditions, retain the subchain with the highest suspicion; When the plurality of sub-chains do not meet the specified condition, the sub-chains are separated to become new second target parent chains.

8. The method according to claim 1, characterized in that The performing neighborhood constraint processing on the second image corresponding to each second target in the high-suspected target chain to obtain a plurality of third images includes: performing data processing on the highly suspected target chain to suppress the complex background of the highly suspected target chain; Filtering the processed high-suspect target chain to remove clutter from the complex background; Performing neighborhood convolution processing on the high-suspected target chain after clutter removal to obtain information about the second target; Calculating each of two adjacent frames of the second image using the information of the second target to obtain a plurality of tracks of the second target; Constraint processing is performed on the tracks of the plurality of second targets to obtain the plurality of third images.

9. The method according to claim 8, characterized in that Counting the number of times the second target appears in the plurality of third images and determining whether the first target is the target to be detected includes: Searching the plurality of third images to detect whether the second target exists in each of the third images, Wherein, when the track of the second target satisfies the constraint condition, the second target exists in the third image; Counting the number of times the second target appears in the plurality of third images and comparing the result with a set threshold; When it is greater than or equal to the threshold, the first target is the target to be detected.

10. A target information continuous generation system, characterized in that: include: The target detection module is used to collect the features of the target to be detected and build a detection model for detecting the target to be detected; a target extraction module, which obtains a sequence of remote sensing images to be processed, wherein the sequence of remote sensing images includes a plurality of consecutive frames of images, determines a first image in which a suspected target to be detected first appears from the sequence of remote sensing images, and extracts a plurality of second images that are consecutive after the first image from the sequence of remote sensing images; a target processing module, configured to perform suspicion degree calculation and neighborhood constraint processing on the plurality of second images to obtain a plurality of third images; The target judgment module performs a threshold comparison on the plurality of the third images to confirm whether the first target that appears for the first time in the first image as the target to be detected is the target to be detected.

Images