A time-series image registration driven spatial weak small moving target detection method and system

Abstract

The present application relates to the technical field of computer image, and provide a kind of time series image registration driven space small moving target detection method and system, including obtaining the time series image short sequence of space small moving target detection;Interframe registration is carried out to time series image short sequence, and registration transformation matrix is obtained, geometric correction is carried out to subsequent frame, and spatial registration is completed;Based on the time series data of completing spatial registration, double labeling is carried out;The time series image short sequence of completing spatial registration is input to the target detection model trained, and detection result is obtained, by interframe registration to time series image short sequence and construct double labeling based on motion sequence frame and coordinate point, so that small moving target can still keep feature alignment and trajectory consistency under the condition of cross-frame background disturbance and low signal-to-noise ratio;Combined with the synergistic structure of encoder and two-stage detection head, the detection and high-precision positioning of weak target such as space debris and microsatellite are realized, and the detection stability and accuracy of space small moving target are improved.

Description

Technical Field

[0001] This invention relates to the field of computer image technology, and in particular to a method and system for detecting spatially weak moving targets driven by temporal image registration. Background Technology

[0002] High-precision detection of small, moving targets is of significant engineering importance and application value in fields such as space situational awareness, on-orbit monitoring of spacecraft, and space debris early warning. Small, moving targets in space typically refer to targets that are dimly lit, small in size, and move rapidly, such as microsatellites, non-cooperative space devices, and space debris. Because these targets exhibit low signal strength during imaging and are often accompanied by complex background noise and illumination changes, traditional single-frame image detection methods struggle to achieve stable and accurate identification.

[0003] Existing spatial target detection methods can be broadly categorized into two types: static detection methods based on single-frame images, which extract targets through photometric thresholding, morphological analysis, or background modeling. However, these methods have limited ability to detect weak signal targets and are prone to generating numerous false detections in high-noise environments. Motion detection methods based on multi-frame images, which detect targets by analyzing brightness changes and motion characteristics over time. However, these methods typically rely on precise image registration, and registration errors can severely impact detection performance when the observation platform exhibits attitude changes or imaging conditions are complex. Furthermore, traditional temporal detection methods often employ single-stage detection structures in target annotation and model design, performing limited modeling of spatial or temporal features, making it difficult to simultaneously consider motion trajectory information and precise target spatial location.

[0004] Despite significant advancements in deep learning technologies for target recognition and motion analysis, existing target detection networks still face numerous challenges when directly applied to temporal spatial imagery. Noise interference and platform jitter in spatial imagery make alignment of multiple frames difficult, and traditional convolutional structures struggle to effectively capture the spatiotemporal correlation features of weak, moving targets. Therefore, designing an algorithm that can fully utilize temporal imagery information and stably detect weak, moving targets in space under low signal-to-noise ratio conditions is a pressing technical challenge in the fields of aerospace remote sensing and space monitoring. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient detection accuracy, difficult image registration, and inaccurate target localization in existing spatial weak moving target detection methods under low signal-to-noise ratio conditions. This invention proposes a temporal image registration-driven spatial weak moving target detection method and system to improve the detection and localization accuracy of weak moving targets in complex backgrounds, especially in low signal-to-noise ratio and high-noise environments, ensuring that targets can be stably and accurately identified and located. This objective can be achieved through the following technical solutions:

[0006] This invention provides a spatial weak moving target detection method driven by temporal image registration, comprising the following steps:

[0007] Acquire short temporal image sequences for detecting spatially weak moving targets;

[0008] Inter-frame registration of short temporal image sequences is performed by dividing each image into multiple patches, selecting the pixel with the highest brightness in each patch as a reference point, calculating the positional offset of the reference point in each frame, obtaining the registration transformation matrix, and performing geometric correction on subsequent frames to complete spatial registration.

[0009] Double annotation is performed on the time-series data that has completed spatial registration. The labeled data is used to train the target detection model. The labeled data includes motion sequence bounding box annotations for modeling the target's motion trajectory and coordinate point annotations for locating the target's position.

[0010] The spatially registered temporal image short sequence is input into the trained target detection model to obtain the detection results of spatially weak moving targets. The target detection model includes an encoder, a first detection head and a second detection head. The first detection head predicts motion sequence boxes based on the spatiotemporal features output by the encoder, and the second detection head locates the target coordinates based on the candidate regions of the motion sequence boxes.

[0011] Furthermore, short temporal image sequences for detecting spatially weak moving targets are obtained, including:

[0012] Temporal image sequences acquired from space imaging devices ;

[0013] Based on the set window length and step size, the time series image sequence is... Divided into several fixed-length time-series image short sequences ,in,

[0014] ;

[0015] ;

[0016] in, For indexing short sequences of time-series images, when hour, For effective time-series image short sequences, For the time-series image sequence, the first... A frame of image acquired at each sampling time. For window length, The step size.

[0017] Furthermore, inter-frame registration is performed on the short temporal image sequence. This involves dividing each image into multiple patches, selecting the pixel with the highest brightness in each patch as a reference point, and calculating the positional offset of the reference point in each frame to obtain the registration transformation matrix. Geometric correction is then performed on subsequent frames to complete spatial registration, including:

[0018] Each image within a short sequence of time-series images is divided into... There are [number] tiles, totaling [number] tiles. Sub-regions;

[0019] In each tile, the pixel with the highest brightness is selected as the reference point to form a set of reference points. ;

[0020] Using the first frame of the short temporal image sequence as the reference frame, the position offset of the corresponding reference point in each frame is calculated to obtain the registration transformation matrix;

[0021] By performing geometric correction on subsequent frames using the registration transformation matrix, spatial registration of each frame within a short sequence of temporal images is achieved.

[0022] Furthermore, dual annotation is performed based on the spatially registered temporal data, including:

[0023] The annotation format for motion sequence boxes is as follows: ,in: Represents the spatial coordinates of the center point of the target sequence. For the corresponding time coordinates; These represent the width, height, and duration of the target sequence, respectively. Indicates the target category;

[0024] The format for labeling coordinate points is as follows: ,in, For spatial coordinates, The time coordinate is used for training the precise location of the target.

[0025] Furthermore, training the object detection model includes:

[0026] Based on the constructed target detection model, a short sequence of temporal images that has undergone spatial registration and double annotation is input for feature extraction and motion trajectory prediction.

[0027] An encoder is used to extract spatiotemporal features from short sequences of input temporal images. The encoder includes a three-dimensional convolutional network or a temporal transformer structure.

[0028] Based on the extracted spatiotemporal features, the motion sequence bounding box is predicted by the first detection head, and candidate regions are generated;

[0029] The second detection head detects the target coordinates within the candidate area output by the first detection head, thus completing the target localization.

[0030] Furthermore, the loss function of the object detection model is a weighted combination of sequence box detection loss, coordinate point regression loss, and classification loss.

[0031] Furthermore, based on the set of reference points constructed from the pixels with the highest brightness in each frame of the image, the registration transformation matrix is ​​obtained by least squares estimation according to the positional offset of the reference points. The registration transformation matrix is ​​either an affine transformation matrix or a perspective transformation matrix.

[0032] Furthermore, the encoder employs a three-dimensional convolutional structure with a convolutional kernel size of [size missing]. It is used to extract spatial and temporal features simultaneously.

[0033] Based on the same inventive concept, this invention provides a spatial weak moving target detection system driven by temporal image registration, employing the spatial weak moving target detection method described above, including:

[0034] The data slicing module is used to acquire short temporal image sequences for detecting spatially weak moving targets;

[0035] The registration module is used to perform inter-frame registration of short sequences of temporal images. It divides each image into multiple patches, selects the pixel with the highest brightness in each patch as a reference point, calculates the position offset of the reference point in each frame, obtains the registration transformation matrix, performs geometric correction on subsequent frames, and completes spatial registration.

[0036] The annotation module is used to perform double annotation based on the spatially registered temporal data, and to obtain the annotation data to train the target detection model. The annotation data includes motion sequence bounding box annotations for modeling the target's motion trajectory, and coordinate point annotations for locating the target's position.

[0037] The detection model module is used to input the short sequence of temporal images that have completed spatial registration into the trained target detection model to obtain the detection results of spatially weak moving targets. The target detection model includes an encoder, a first detection head and a second detection head. The first detection head predicts motion sequence boxes based on the spatiotemporal features output by the encoder, and the second detection head locates the target coordinates based on the candidate regions of the motion sequence boxes.

[0038] Furthermore, the data slicing module also includes a data processing unit for processing time-series image sequences acquired from the space imaging device. Based on the set window length and step size, the time series image sequence is... Divided into several fixed-length time-series image short sequences ,in,

[0039] ;

[0040] ;

[0041] in, For indexing short sequences of time-series images, when hour, For effective time-series image short sequences, For the time-series image sequence, the first... A frame of image acquired at each sampling time. For window length, The step size.

[0042] Compared with the prior art, the present invention has at least one of the following technical advantages:

[0043] This invention enables weak moving targets to maintain feature alignment and trajectory consistency even under cross-frame background disturbances and low signal-to-noise ratio conditions by performing inter-frame registration on short sequences of temporal images and constructing dual annotations based on motion sequence bounding boxes and coordinate points. Simultaneously, by combining the collaborative structure of the encoder and the two-stage detection head, the first detection head first extracts the target motion sequence bounding box, and then the second detection head precisely regresses the coordinate points in the candidate region, achieving robust detection and high-precision positioning of weak targets such as space debris and microsatellites. This significantly improves the detection stability and accuracy of weak moving targets in space, and has important value for aerospace remote sensing and space security applications. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below:

[0045] Figure 1 This is a flowchart of the spatial weak moving target detection method driven by time-series image registration according to the present invention;

[0046] Figure 2 The temporal spatial plots and their difference diagrams provided for experiments in embodiments of the present invention, wherein, Figure 2 (a), (b), and (c) in the image are three consecutive spatial images from a short temporal image sequence. Figure 2 (d) in the image is a difference map generated from three consecutive spatial images;

[0047] Figure 3 The above are the time-series spatial plots and their differences that have been registered in this embodiment of the invention. Figure 3 In the image, (a), (b), and (c) are three consecutive spatial images obtained after registration processing in a short temporal image sequence. Figure 3 (d) in the figure is the difference map generated based on the spatial images after registration of three consecutive frames;

[0048] Figure 4 This refers to the target motion frame detection result in this embodiment of the invention;

[0049] Figure 5 The target location detection result in this embodiment of the invention is shown below. Figure 5 (a), (b), and (c) in the figure represent the target location detection results in three consecutive frames. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] First Embodiment

[0052] In missions such as space situational awareness, spacecraft on-orbit health monitoring, and space debris early warning, researchers have long relied on time-series images acquired by optical payloads to monitor the orbital environment. However, in real-world observation scenarios, small moving targets such as microsatellites, non-cooperative space devices, and micro-debris are often only 1 to 3 pixels in size, with brightness far lower than background stars and noise. Engineers have generally reported that if the platform's attitude fluctuates slightly, the lighting changes drastically, or the background star field is complex, the target can be completely obscured in a single frame of the image, making it very difficult to identify directly from the original image.

[0053] To improve detection capabilities, existing technologies mainly employ two types of methods. The first type is traditional single-frame static detection methods, such as photometric thresholding, background modeling, or morphological analysis. These methods are effective when the target signal is strong, but in scenarios with weak targets, the brightness of a single frame is almost equal to the noise itself, making it difficult to set the threshold and resulting in extremely high false detection rates, generally leading to poor performance in engineering practice. The second type of method attempts to utilize motion information between multiple frames to perform differential analysis, temporal filtering, or motion estimation based on optical flow in the image sequence, hoping to enhance the detection of weak targets by leveraging the continuity of the temporal dimension. However, these methods often rely on precise inter-frame registration, while actual observation platforms inevitably experience attitude perturbations and field-of-view drift, resulting in the differential image being filled with a large number of pseudo-motion signals caused by jitter. The motion amplitude of the weak target itself is extremely small, causing its trajectory features to be submerged in jitter noise, thus making reliable detection impossible.

[0054] Even with the rise of deep learning technology, attempts have begun to apply mainstream video detection networks directly to spatial temporal images, but these models often struggle to adapt to complex starry backgrounds and low signal-to-noise ratio conditions. The features of small targets are extremely insignificant, making it difficult for deep networks to establish stable spatiotemporal relationships. Furthermore, existing methods generally employ single-stage detection structures, which fail to simultaneously represent the target's trajectory and its precise spatial location in terms of annotation strategies and model design. As a result, although deep models theoretically possess learning capabilities, their detection performance in the domain of small spatial targets remains unstable, with high false positive rates and extreme sensitivity to platform fluctuations.

[0055] Therefore, how to construct a weak moving target detection method that can effectively utilize multi-frame image information and achieve stable registration, accurate tracking and fine positioning under conditions of noise dominance, frequent attitude perturbations and no visibility in a single frame remains a core problem that needs to be solved by existing technologies.

[0056] Faced with the aforementioned technical bottlenecks, the inventors, after in-depth analysis, noticed a core principle: although weak moving targets are almost completely submerged by noise in a single frame of imagery, they often exhibit continuous, stable, and consistent weak motion characteristics over time. If a reliable inter-frame spatial alignment mechanism can be constructed in temporal images, allowing the effective signals of the same target to be consistently superimposed across multiple frames, the signal-to-noise ratio can be significantly improved, breaking through the limits of single-frame detection. Simultaneously, the inventors also discovered that traditional feature point registration (such as SIFT, ORB) or global cross-correlation registration is easily affected by noise, star points, or platform jitter in scenes with weak targets, leading to failure. While brightness peaks in images are often considered noise, using the pixel with the highest brightness as a local reference point under block constraints can create a sparse but stable brightness peak reference structure, thus providing new possibilities for temporal registration under low signal-to-noise ratio conditions.

[0057] Based on this key observation, the inventors further realized that relying solely on traditional single-layer target bounding box annotations is insufficient to simultaneously express the "temporal range of existence" and "precise spatial location" of weak moving targets. Therefore, a labeling system capable of simultaneously encompassing the target's spatiotemporal motion range and fine-grained positioning information is urgently needed to support higher-level learning capabilities. As their research progressed, the inventors proposed a two-layer annotation method using motion sequence bounding boxes plus coordinate points. This allows the model to simultaneously learn the target's motion envelope in the temporal dimension and its precise center point in the spatial dimension, improving the reliability and accuracy of weak target detection.

[0058] Building upon this foundation, the inventors ultimately constructed a novel temporal image-driven detection framework: firstly, robust inter-frame registration is achieved through sparse brightness peak reference points; subsequently, a two-stage detection network is built based on a dual-layer annotation system. The first stage is responsible for identifying the target's motion range in the temporal dimension, while the second stage performs refined coordinate point detection within the candidate region, thus simultaneously solving the long-standing challenges of whether weak targets can be detected and whether they can be accurately located. Therefore, this invention achieves stable identification and high-precision positioning of weak moving targets in complex spatial backgrounds, providing a new technical approach for the fields of aerospace remote sensing and space security. Specific implementation methods are as follows:

[0059] like Figure 1 As shown, this invention provides a spatial weak moving target detection method driven by temporal image registration, comprising the following steps:

[0060] Acquire short temporal image sequences for detecting spatially weak moving targets;

[0061] Inter-frame registration of short temporal image sequences is performed by dividing each image into multiple patches, selecting the pixel with the highest brightness in each patch as a reference point, calculating the positional offset of the reference point in each frame, obtaining the registration transformation matrix, and performing geometric correction on subsequent frames to complete spatial registration.

[0062] Double annotation is performed on the time-series data that has completed spatial registration. The labeled data is used to train the target detection model. The labeled data includes motion sequence bounding box annotations for modeling the target's motion trajectory and coordinate point annotations for locating the target's position.

[0063] A short sequence of temporal images with completed spatial registration is input into a converged target detection model. The model sequentially outputs the motion bounding boxes and coordinate points of the targets, obtaining the detection results of small, moving targets in space, thus achieving automatic detection of small, moving targets in space. This method can effectively identify and locate small-scale spatial targets (such as microsatellites and space debris) under conditions of low brightness and low signal-to-noise ratio. The target detection model includes an encoder, a first detection head, and a second detection head. The first detection head predicts motion bounding boxes based on the spatiotemporal features output by the encoder, and the second detection head locates the target coordinate points based on candidate regions of the motion bounding boxes.

[0064] Furthermore, short temporal image sequences for detecting spatially weak moving targets are obtained, including:

[0065] Temporal image sequences acquired from space imaging devices ;

[0066] Based on the set window length and step size, the time series image sequence is... Divided into several fixed-length time-series image short sequences ,in,

[0067] ;

[0068] ;

[0069] in, For indexing short sequences of time-series images, when hour, For effective time-series image short sequences, For the time-series image sequence, the first... A frame of image acquired at each sampling time. For window length, The step size.

[0070] Furthermore, inter-frame registration is performed on the short temporal image sequence. This involves dividing each image into multiple patches, selecting the pixel with the highest brightness in each patch as a reference point, and calculating the positional offset of the reference point in each frame to obtain the registration transformation matrix. Geometric correction is then performed on subsequent frames to complete spatial registration, including:

[0071] Each image within a short sequence of time-series images is divided into... There are [number] tiles, totaling [number] tiles. Sub-regions;

[0072] In each tile, the pixel with the highest brightness is selected as the reference point to form a set of reference points. ;

[0073] Using the first frame of a short temporal image sequence as the reference frame, the registration transformation matrix is ​​obtained by calculating the position offset of the reference points in adjacent frames;

[0074] The registration transformation matrix is ​​used to perform affine or perspective correction on subsequent frames to ensure that each frame in the short sequence is aligned with the reference frame.

[0075] Furthermore, dual annotation is performed based on the spatially registered temporal data, including:

[0076] The annotation format for motion sequence boxes is as follows: ,in: Represents the spatial coordinates of the center point of the target sequence. For the corresponding time coordinates; These represent the width, height, and duration of the target sequence, respectively. Indicates the target category;

[0077] The format for labeling coordinate points is as follows: ,in, For spatial coordinates, The time coordinate is used to accurately locate the center point coordinates of the target at a specific moment.

[0078] Furthermore, training the object detection model includes:

[0079] Based on the constructed target detection model, a short sequence of temporal images that has undergone spatial registration and double annotation is input for feature extraction and motion trajectory prediction.

[0080] An encoder is used to extract spatiotemporal features from short sequences of input temporal images. The encoder includes a three-dimensional convolutional network (3D-CNN) or a temporal transformer structure to capture both spatial and temporal features simultaneously.

[0081] Based on the extracted spatiotemporal features, the motion sequence bounding box is predicted by the first detection head, and candidate regions are generated;

[0082] Specifically, the first detection head is a motion sequence bounding box detection head. The input is the feature map extracted by the encoder, and the output is the motion sequence bounding box of the target, so as to achieve coarse localization of the candidate target.

[0083] The second detection head detects the target coordinates within the candidate area output by the first detection head, thus completing the target localization.

[0084] Specifically, the second detection head is a coordinate point detection head. It takes the candidate results output by the first detection head as input and predicts the specific coordinate points of the target within the corresponding area, thereby achieving refined detection of the target location.

[0085] Furthermore, the loss function of the object detection model is a weighted combination of the sequence box detection loss, the coordinate point regression loss, and the classification loss.

[0086] Specifically, the loss function design is as follows: a multi-task joint loss function is adopted, defined as:

[0087] ;

[0088] in, For sequence frame detection loss, For coordinate point regression loss, For classifying losses, , , These are the weighting coefficients.

[0089] During training, the model parameters are optimized by evaluating the performance on the validation set until the model converges.

[0090] Preferably, to improve the robustness of the model, data augmentation methods such as random temporal pruning, spatial translation, brightness perturbation, and rotation enhancement can be used to augment the samples.

[0091] Furthermore, based on the set of reference points constructed from the pixels with the highest brightness in each frame of the image, the registration transformation matrix is ​​obtained by least squares estimation according to the positional offset of the reference points. The registration transformation matrix is ​​either an affine transformation matrix or a perspective transformation matrix.

[0092] Furthermore, the encoder employs a three-dimensional convolutional structure with a convolutional kernel size of [size missing]. It is used to extract both spatial and temporal features simultaneously.

[0093] Specifically, the model's encoder may include multiple spatiotemporal convolutional layers, with the kernel size set to [value missing]. Each convolutional layer is followed by a ReLU activation function to enhance nonlinear feature extraction capabilities. The motion sequence boxes and coordinate point results output by the model can be optimized for detection accuracy using thresholding and non-maximum suppression (NMS) algorithms, respectively.

[0094] The invention will be further illustrated below through specific experiments:

[0095] Figure 2 This is a schematic diagram of a temporal spatial image for experimental verification provided in an embodiment of the present invention. For ease of demonstration, this image is taken only from a local area of ​​the original observation image, and the temporal sequence length is set to 3. Figure 2 (a), (b), and (c) in the figure correspond to three consecutively acquired spatial images. It can be seen that the original temporal image contains a large number of star points, noise specks, and background brightness fluctuations; the brightness of weak moving targets is extremely low, and their shapes are blurred. Yellow, cyan, purple, and green ellipses in the figure mark candidate points of weak targets located in different regions, respectively. Figure 2 (d) in the middle is Figure 2 The difference diagrams of frames (a), (b), and (c) show that weak targets that have shifted over time are highlighted by increased brightness or strip-like structures, thereby suppressing background stars and enhancing real moving targets. It can be seen that there is a slight overall shift between different frames, providing a typical application scenario for subsequent inter-frame registration and temporal detection processes. Figure 3 These are the time-series spatial map blocks that have been registered in this embodiment of the invention. Figure 3 In the figures, (a), (b), and (c) are the aligned images after processing the three original images using the brightness reference point-driven registration method described in this invention. Yellow, cyan, purple, and green ellipses are used to mark weak target candidate points located in different regions. Figure 3 (d) in the middle is Figure 3 The difference diagrams of frames (a), (b), and (c) show that the registered images maintain a high degree of consistency in star distribution, background structure, and noise texture, and the inter-frame drift is significantly reduced, providing a stable and consistent input for subsequent temporal feature extraction, motion sequence box detection, and coordinate point localization. Figure 4This is the target motion bounding box detection result in an embodiment of the present invention; the different colored markers in the figure correspond to the target motion trajectory range detected by the model in a short sequence of time-series images, where the red, green, and blue markers are the corresponding positions of the target in the first frame, second frame, and third frame, respectively. Figure 4 As can be seen, there are four distinct target trajectories in the current image.

[0096] Because small moving targets have extremely low brightness in a single frame, they are difficult to identify directly with the naked eye. This invention first performs inter-frame spatial registration, and then, based on temporal feature extraction and motion sequence bounding box prediction, enables the subtle movement of the target across multiple frames to be revealed. Its motion trajectory is presented in the image as a short line segment structure composed of several colored dots. The detection results show that this invention can accurately identify the presence area and movement range of small moving targets under complex starry backgrounds and low signal-to-noise ratio conditions, providing reliable candidate regions for subsequent coordinate-level fine localization. Figure 5 This is the target location detection result in an embodiment of the present invention. Figure 5 Figures (a), (b), and (c) show the target detection results obtained after processing three consecutive frames of sky area observation images. The locations marked with yellow circles in the figures are suspected target points identified by the detection algorithm of this invention. Figure 5 As shown, even under conditions of strong background noise, dense star points, and uneven brightness distribution, the weak target detection method of this invention can still stably extract the target location from the complex background, and the detection results in consecutive frames of images show good consistency. In particular, for regions on the right side of the image with brightness gradient changes and fixed stripe interference, the method of this invention still maintains high detection accuracy, with no obvious false positives or false negatives. Therefore, the detection method of this invention can reliably detect weak targets in the sky under conditions of weak signals, low signal-to-noise ratio, and non-uniform background, verifying the effectiveness of the algorithm in actual observation scenarios.

[0097] Second Embodiment

[0098] Based on the same inventive concept, this invention provides a spatial weak moving target detection system driven by temporal image registration, employing the spatial weak moving target detection method described above, including:

[0099] The data slicing module is used to acquire or receive short temporal image sequences for detecting spatially weak moving targets, and to slice the input image sequences according to the window length and step size.

[0100] The registration module is used to perform inter-frame registration of short sequences of temporal images. It divides each image into multiple patches, selects the pixel with the highest brightness in each patch as a reference point, calculates the position offset of the reference point in each frame, obtains the registration transformation matrix, performs geometric correction on subsequent frames, and completes spatial registration.

[0101] The annotation module is used to perform double annotation based on the spatially registered temporal data, and to obtain the annotation data to train the target detection model. The annotation data includes motion sequence bounding box annotations for modeling the target's motion trajectory, and coordinate point annotations for locating the target's position.

[0102] The detection model module is used to execute the above two-stage detection model, perform target detection on the input short temporal images, input the spatially registered short sequence of temporal images into the trained target detection model, and obtain the detection results of spatially weak moving targets. The target detection model includes an encoder, a first detection head and a second detection head. The first detection head predicts the motion sequence box of the target based on the spatiotemporal features output by the encoder, and the second detection head locates the target based on the candidate region of the motion sequence box and determines the target coordinate point.

[0103] Furthermore, the data slicing module also includes a data processing unit for processing time-series image sequences acquired from the space imaging device. Based on the set window length and step size, the time series image sequence is... Divided into several fixed-length time-series image short sequences ,in,

[0104] ;

[0105] ;

[0106] in, For indexing short sequences of time-series images, when hour, For effective time-series image short sequences, For the time-series image sequence, the first... A frame of image acquired at each sampling time. For window length, The step size.

[0107] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A spatial weak moving target detection method driven by temporal image registration, characterized in that, Includes the following steps: Acquire short temporal image sequences for detecting spatially weak moving targets; Inter-frame registration is performed on the short sequence of temporal images by dividing each image into multiple blocks, selecting the pixel with the highest brightness in each block as a reference point, calculating the positional offset of the reference point in each frame, obtaining the registration transformation matrix, and performing geometric correction on subsequent frames to complete spatial registration. Based on the time-series data that has completed the spatial registration, double annotation is performed to obtain the annotation data for training the target detection model. The annotation data includes motion sequence box annotations for modeling the target's motion trajectory and coordinate point annotations for locating the target's position. The short sequence of temporal images that has completed spatial registration is input into the trained target detection model to obtain the detection results of spatially weak moving targets. The target detection model includes an encoder, a first detection head and a second detection head. The first detection head predicts the motion sequence box of the target based on the spatiotemporal features output by the encoder, and the second detection head locates the target coordinate point based on the candidate region of the motion sequence box.

2. The spatial weak moving target detection method driven by temporal image registration according to claim 1, characterized in that, Acquire short temporal image sequences for detecting spatially weak moving targets, including: Temporal image sequences acquired from space imaging devices ; Based on the set window length and step size, the time-series image sequence is... The time-series images are divided into several short sequences of fixed length, wherein, ； ； in, For the index of the short sequence of time-series images, when hour, For effective time-series image short sequences, The first in the time-series image sequence A frame of image acquired at each sampling time. The length of the window. Let be the step size.

3. The spatial weak moving target detection method driven by temporal image registration according to claim 1, characterized in that, Inter-frame registration is performed on the short sequence of temporal images by dividing each image into multiple patches, selecting the pixel with the highest brightness in each patch as a reference point, calculating the positional offset of the reference point in each frame to obtain the registration transformation matrix, and performing geometric correction on subsequent frames to complete spatial registration, including: Each image in the short sequence of time-series images is divided into... The total number of the aforementioned blocks is [number]. Sub-regions; In each of the aforementioned blocks, the pixel with the highest brightness is selected as the reference point, forming a reference point set. ; Using the first frame of the short temporal image sequence as the reference frame, the position offset of each frame corresponding to the reference point is calculated to obtain the registration transformation matrix; The geometric correction is performed on the subsequent frames using the registration transformation matrix to achieve spatial registration of the images in each frame of the short temporal image sequence.

4. The spatial weak moving target detection method driven by temporal image registration according to claim 1, characterized in that, Double annotation is performed based on the temporal data that has completed the spatial registration, including: The annotation format for the motion sequence frame is as follows: ,in: Represents the spatial coordinates of the center point of the target sequence. For the corresponding time coordinates; These represent the width, height, and duration of the target sequence, respectively. Indicates the target category; The format for labeling the coordinate points is as follows: ,in, The spatial coordinates are... The time coordinates are used for training the precise location of the target.

5. The spatial weak moving target detection method driven by temporal image registration according to claim 1, characterized in that, Training the target detection model includes: Based on the constructed target detection model, the short temporal image sequence that has undergone spatial registration and dual annotation is input to perform feature extraction and motion trajectory prediction; The encoder is used to extract the spatiotemporal features from the input short sequence of temporal images, wherein the encoder includes a three-dimensional convolutional network or a temporal transformer structure; Based on the extracted spatiotemporal features, the motion sequence box is predicted by the first detection head, and the candidate region is generated; The second detection head detects the target coordinates within the candidate area output by the first detection head, thus completing the target localization.

6. The spatial weak moving target detection method driven by temporal image registration according to claim 1, characterized in that, The loss function of the target detection model is a weighted combination of sequence box detection loss, coordinate point regression loss, and classification loss.

7. The spatial weak moving target detection method driven by temporal image registration according to claim 3, characterized in that, Based on the reference point set constructed from the pixel with the highest brightness in each frame of the image, the registration transformation matrix is ​​obtained by least squares estimation according to the positional offset of the reference points. The registration transformation matrix is ​​either an affine transformation matrix or a perspective transformation matrix.

8. The spatial weak moving target detection method driven by temporal image registration according to claim 7, characterized in that, The encoder employs a three-dimensional convolutional structure to simultaneously extract spatial and temporal features.

9. A spatial weak moving target detection system driven by temporal image registration, employing the spatial weak moving target detection method driven by temporal image registration as described in any one of claims 1 to 8, characterized in that, include: The data slicing module is used to acquire short temporal image sequences for detecting spatially weak moving targets; The registration module is used to perform inter-frame registration on the short sequence of temporal images. It divides each image into multiple blocks, selects the pixel with the highest brightness in each block as a reference point, calculates the position offset of the reference point in each frame, obtains the registration transformation matrix, performs geometric correction on subsequent frames, and completes spatial registration. The annotation module is used to perform double annotation based on the temporal data that has completed the spatial registration, and to obtain the annotation data to train the target detection model. The annotation data includes motion sequence box annotations for modeling the target motion trajectory and coordinate point annotations for locating the target position. The detection model module is used to input the short sequence of temporal images that have completed the spatial registration into the trained target detection model to obtain the detection results of spatially weak moving targets. The target detection model includes an encoder, a first detection head and a second detection head. The first detection head predicts the motion sequence box based on the spatiotemporal features output by the encoder, and the second detection head locates the target coordinate point based on the candidate region of the motion sequence box.

10. The spatial weak moving target detection system driven by temporal image registration according to claim 9, characterized in that, The data slicing module also includes a data processing unit for processing time-series image sequences acquired from the space imaging device. Based on the set window length and step size, the time-series image sequence is... The time-series images are divided into several short sequences of fixed length, wherein, ； ； in, For the index of the short sequence of time-series images, when hour, For effective time-series image short sequences, The first in the time-series image sequence A frame of image acquired at each sampling time. The length of the window. Let be the step size.

Images