Dense and small target detection method and system based on time series analysis

Abstract

The present invention belongs to the field of moving target detection, and relates to a method and system for detecting dense and weak targets based on time series analysis. The method comprises: preprocessing the received image sequence to obtain a pixel time series intensity signal corresponding to each pixel; using time series analysis technology to process each pixel time series intensity signal, repeating this step until all reconstructed pixel time series intensity signals containing only target signals are obtained; obtaining signal data corresponding to each frame in all reconstructed signals according to the number of frames, and reconstructing the detection image corresponding to the frame; using density clustering to process the detection image, clustering discrete target pixel points into targets, and eliminating false alarm points; after obtaining all pixels corresponding to the target, using the matrix external bounding box method to obtain the bounding box of the target, and adding the bounding box to the original image to obtain the detection result of the frame, and repeating this step until the detection results of all frames are obtained.

Description

Technical Field

[0001] The present invention belongs to the field of moving target detection, and in particular relates to a method and system for detecting dense and dim targets based on time series analysis. Background Art

[0002] Moving target detection has important applications in military defense, traffic management, ocean monitoring and other fields, and is also a research hotspot at home and abroad. In actual surveillance and early warning systems, in order to detect targets as quickly and early as possible, the detection distance is required to be as far as possible and the detection field of view as large as possible. However, under long-distance and large-field-of-view imaging conditions, the moving target only occupies a few pixels after imaging, the target energy becomes weak, and the airspace features such as shape, texture, and size are lost, presenting as a weak target. At the same time, with the rapid development of detection systems and detection targets, in some special detection scenarios, there may be a large number of weak targets in the field of view, and the targets are dense in the airspace, which is of great significance to the effective detection of dense and weak targets in the detection field of view.

[0003] Moving target detection methods can be divided into model-driven detection methods and data-driven detection methods. Model-driven detection methods manually design detectors based on the feature differences between the target and the background, while data-driven detection methods achieve detection by designing neural networks and allowing them to autonomously learn target features from a large amount of data. In most fields, the performance of data-driven detection methods represented by deep learning is far ahead of model-driven detection methods. However, in the detection of weak targets, due to the weak target energy and lack of spatial features, data-driven detection methods are difficult to effectively learn target features. At the same time, dense targets will be confused with clutter in the background, reducing the reliability of the detection results of data-driven methods. Therefore, model-driven detection methods still occupy the mainstream in the field of weak target detection due to their interpretability, high reliability and stable detection performance.

[0004] In essence, model-driven target detection methods mainly use two types of information: one is visual features, namely the appearance features of the target such as shape, texture, size, and the statistical characteristics of the target and the background such as contrast, variance, and mean; the other is the motion information of the target, including motion direction, motion state, etc. According to the different information used, the detection methods can be divided into single-frame detection methods and multi-frame detection methods. Single-frame detection methods include filtering-based methods, methods based on the human visual system, and methods based on low-rank matrix sparse decomposition. In general, since small and weak targets have no obvious spatial characteristics and dense targets do not meet many constraints in single-frame detection methods, single-frame detection methods are less effective. The existing multi-frame detection method only introduces time domain information for assistance on the basis of spatial information, and does not fully explore the motion characteristics of moving targets in the time domain. The detection performance is extremely limited for dense and small targets. Summary of the invention

[0005] In view of the problem that existing methods cannot effectively detect dense and small targets, the purpose of the present invention is to overcome the above-mentioned defects of the prior art and propose a dense and small target detection method based on time series analysis.

[0006] In order to achieve the above object, the present invention proposes a dense dim small target detection method based on time series analysis, comprising:

[0007] Step A: preprocess the received image sequence to obtain the pixel time series intensity signal corresponding to each pixel;

[0008] Step B: using time series analysis technology to process each pixel time series intensity signal, repeating this step until all reconstructed pixel time series intensity signals containing only the target signal are obtained;

[0009] Step C: obtaining the signal data corresponding to each frame in all reconstructed signals according to the frame number, and reconstructing the detection image corresponding to the frame;

[0010] Step D: Use density clustering to process the detection image, cluster discrete target pixel points into targets, and eliminate false alarm points;

[0011] Step E: After obtaining all the pixels corresponding to the target, the matrix external bounding box method is used to obtain the bounding box of the target, and the bounding box is added to the original image to obtain the detection result of the frame. This step is repeated until the detection results of all frames are obtained.

[0012] Preferably, step A comprises:

[0013] Step A1: Obtain the pixel time series intensity signal corresponding to each pixel according to the coordinates;

[0014] Step A2: Use the corrected mean and standard deviation of the pixel time series intensity signal as the mean and standard deviation of the filling signal to generate the filling signal according to the normal distribution

[0015]

[0016] Among them, α and β are artificially set hyperparameters used to offset the impact of the target signal on the statistical characteristics of the pixel time series intensity signal; μ and σ represent the mean and standard deviation of the original pixel time series intensity signal. and is the corrected mean and standard deviation, N() indicates normal distribution;

[0017] Get the pixel time series intensity signal after filling for:

[0018]

[0019] δ represents the length of the padding signal, and Concate represents the signal concatenation operation.

[0020] Step A3: Gaussian filtering is performed on the pixel time series intensity signal. The filtered pixel time series intensity signal X G (t) is:

[0021]

[0022] σ g is the standard deviation of Gaussian filtering, which is used to represent the pixel time series intensity signal after filtering;

[0023] Step A4: Filter the pixel time series intensity signal X G (t) Standardize it to satisfy the standard normal distribution:

[0024]

[0025] Among them, μ G and σ G Represents the mean and standard deviation of the pixel time series intensity signal after Gaussian filtering.

[0026] Preferably, the step B comprises:

[0027] Step B1: Initialize hyperparameters, including target signal width W, local standard deviation threshold θ σ The maximum difference between the two peaks and the standard deviation D p ;

[0028] Step B2: Perform the following steps for each pixel time series intensity signal input:

[0029] Step B3: Using the target signal width W as the sliding window, calculate the local standard deviation within the window to form a local standard deviation signal S σ (k):

[0030]

[0031] Among them, μw represents the mean value in the sliding window, k represents the element index of the local standard deviation signal, and t represents the pixel time series intensity signal X G (t) element index;

[0032] Step B4: local standard deviation signal S σ (k) Perform element-by-element determination: If S σ (k) is greater than the set standard deviation threshold θσ, go to step B5; otherwise, traverse the next element, that is, add 1 to k until the element-by-element determination is completed;

[0033] Step B5: Perform extreme point detection and peak analysis: If S σ There are two maximum points k1, k3 and one minimum point k2 in the local area of ​​(k), and |S σ (k1)-S σ (k3)|<D p , where k1, k2, k3 are the element indices corresponding to the extreme points, D p is the maximum difference between the two peaks of the standard deviation; if there is a target signal in the local area, go to step B6; otherwise, add 1 to k and go to step B4;

[0034] Step B6: The start and end indexes of the target signal are the maximum point k, and the element index k4 corresponding to the foot of the second peak of the bimodal pattern in the local standard deviation signal; the elements between k and k4 in the original pixel time series intensity signal are retained, thereby realizing the pixel time series intensity signal X G (t), and obtain the reconstructed signal S T (t):

[0035]

[0036] Traverse the next element, that is, add 1 to k, and go to step B4; iterate the elements until the processing of the pixel time series intensity signal is completed.

[0037] Preferably, the step C comprises:

[0038] Step C1: extracting data at a specified frame of all pixel time series intensity signals frame by frame;

[0039] Step C2: reconstruct the data of all pixel time series intensity signals at the specified frame into a two-dimensional matrix I(k) according to the coordinates:

[0040]

[0041] Where M represents the number of rows of the image, N represents the number of columns of the image, and M*N represents the number of pixels in the image, which is also the number of pixel time series intensity signals; represents the first pixel time series intensity signal, Represents the second pixel time series intensity signal, represents the last pixel time series intensity signal, K represents the Kth frame, and Reshape() represents the reconstruction of a one-dimensional vector into a two-dimensional matrix;

[0042] Step C3: Normalize the two-dimensional matrix to between 0 and 255 to obtain the detection image:

[0043]

[0044] Among them, Max() and Min() represent the maximum value and minimum value respectively.

[0045] Preferably, the step D comprises:

[0046] Step D1: Initialize the density clustering hyperparameters, which include the neighborhood radius ε, the minimum number of points P within the neighborhood radius M ;

[0047] Step D2: Starting from any pixel p with non-zero intensity value, mark it as “visited” and determine whether the neighborhood of p has at least P M If there are non-zero intensity pixels, it is judged as yes, then p is the core pixel, a new target cluster C is created for p, and all non-zero intensity pixels in the neighborhood of p are placed in the candidate target cluster G; otherwise, it is marked as a noise point;

[0048] Step D3: Iteratively add the non-zero intensity pixels in the candidate target cluster G that do not belong to other target clusters to the target cluster C. For the non-zero intensity pixel p' marked as "unvisited" in G, mark it as "visited", and check the neighborhood of the non-zero intensity pixel p'. If p' is a core pixel, add all the non-zero intensity pixels in the neighborhood of p' to the candidate target cluster G. Continue to add non-zero intensity pixels to the target cluster C until the target cluster C cannot be expanded, that is, the candidate target cluster G is empty, and the target cluster C is completely generated.

[0049] Step D4: Randomly select the next non-zero intensity pixel that has not been visited from the remaining non-zero intensity pixels, and go to step D3 until all non-zero intensity pixels are visited, and then end this step.

[0050] Preferably, the step E comprises:

[0051] Step E1: Calculate the target centroid based on all pixels corresponding to the target

[0052]

[0053] Among them, (x i ,y i ) is the pixel coordinate, n is the number of pixels in the target pixel set;

[0054] Step E2: Based on the target pixel set, construct the target covariance matrix Cov:

[0055]

[0056] Among them, i represents the sequence number of the i-th pixel in the target pixel set, and T represents transposition;

[0057] Step E3: performing eigenvalue decomposition on the target covariance matrix to determine the principal component direction;

[0058] Step E4: Calculate the width and height of the bounding box according to the result of eigenvalue decomposition;

[0059] Step E5: Using the centroid and the principal component directions, calculate the coordinates of the four corner points of the bounding box;

[0060] Step E6: Draw the calculated bounding box corner points onto the original image;

[0061] Step E7: Go to step E1 until bounding boxes are generated for all detected objects and added to the original image to form a complete detection result.

[0062] On the other hand, the present invention provides a dense dim small target detection system based on time series analysis, comprising:

[0063] A preprocessing module is used to preprocess the received image sequence to obtain a pixel time-series intensity signal corresponding to each pixel;

[0064] A pixel time series intensity signal reconstruction module is used to process each pixel time series intensity signal using a time series analysis technique, and repeat this step until all reconstructed pixel time series intensity signals containing only the target signal are obtained;

[0065] A detection image reconstruction module is used to obtain signal data corresponding to each frame of all reconstructed signals according to the frame number, and reconstruct the detection image corresponding to the frame;

[0066] The dense target segmentation module is used to process the detection image using density clustering, cluster discrete target pixel points into targets, and eliminate false alarm points;

[0067] The target bounding box acquisition module is used to obtain all pixels corresponding to the target, obtain the target bounding box by using the matrix external bounding box method, and add the bounding box to the original image to obtain the detection result of the frame. The process is repeated until the detection results of all frames are obtained.

[0068] Compared with the prior art, the advantages of the present invention are:

[0069] 1. Aiming at the problem of dense and small target detection, the present invention takes the pixel time series intensity signal formed by a single pixel as the processing object, uses the time series analysis technology to analyze the statistical characteristic difference between the target signal and the background and noise signals, and suppresses the background signal and retains the target signal according to the difference, reconstructs the pixel time series intensity signal, and realizes the effective detection of dense and small targets;

[0070] 2. Aiming at the problem of segmenting dense targets in the reconstructed detection image, the present invention uses a density clustering method to cluster the pixels corresponding to the target, obtain all the pixels corresponding to each target, and solve the segmentation problem between dense targets. At the same time, the use of the density clustering method can also effectively suppress discrete false alarm points in the detection results, greatly reducing the false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is the overall detection idea diagram of the dense dim small target detection method based on time series analysis of the present invention;

[0072] Figure 2 It is a flow chart of the dense dim small target detection method based on time series analysis of the present invention. DETAILED DESCRIPTION

[0073] The present invention starts with the time domain motion information of dense and small targets, uses time series analysis technology to convert the dense and small target detection problem in the spatial domain into the target signal reconstruction problem in the time domain, and inverts the detection results based on the reconstructed pixel signal to achieve high-performance detection of dense and small targets. It includes the following steps:

[0074] Step A: After obtaining the image sequence, the image sequence is preprocessed to obtain the pixel time-series intensity signal corresponding to each pixel.

[0075] Step B: Use time series analysis technology to process each pixel time series intensity signal to obtain the reconstructed pixel time series intensity signal. The reconstructed signal will not contain background and noise signals, but only target signals formed by dense and weak target movements.

[0076] Step C: After processing all pixel time series intensity signals, the signal data corresponding to each frame in all reconstructed signals are obtained according to the frame number, and the detection image corresponding to the frame is reconstructed according to these data to obtain the detection image of each frame.

[0077] Step D: Use density clustering to process the detection image, cluster discrete target pixels into targets, and eliminate false alarm points.

[0078] Step E: After obtaining all pixels corresponding to the target, use the matrix external bounding box method to obtain the target's bounding box, and add the bounding box to the original image to obtain the detection result of the frame. Repeat the operation to obtain the detection results of all frames.

[0079] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0080] Example 1

[0081] like Figure 1As shown, Example 1 of the present invention provides a dense dim target detection method based on time series analysis. The scheme of the present invention consists of five parts: first, the input image sequence is preprocessed to obtain the pixel time series intensity signal corresponding to each pixel. Secondly, for all pixel time series intensity signals, the time series analysis technology is used to suppress the background and noise signals, retain the target signal, and reconstruct the pixel time series intensity signal. After obtaining the reconstructed signal of all pixel time series intensity signals, the signal data corresponding to each frame is obtained according to the number of frames, and the signal data is reconstructed into a detection image. Subsequently, the density clustering method is used to cluster and segment the targets in the detection image to obtain all the pixels corresponding to each target. Finally, the matrix external connection method is used to obtain the target external bounding box, and it is added to the original image to obtain the detection result.

[0082] Step 1: Image sequence preprocessing

[0083] Step A: Preprocess the input image sequence to obtain the pixel time-series intensity signal corresponding to each pixel. The specific process includes:

[0084] Step A1: Obtain the pixel time series intensity signal corresponding to each pixel according to the coordinates.

[0085] Step A2: Fill the pixel time series intensity signal. In order to avoid the problem of missed detection caused by insufficient target motion information at the beginning and end of the image sequence, it is necessary to fill the data before and after the pixel time series intensity signal starts. The specific filling method is: use the corrected mean and standard deviation of the pixel time series intensity signal as the mean and standard deviation of the filling signal, and generate the filling signal according to the normal distribution. The formula is:

[0086]

[0087] α and β are artificially set hyperparameters to offset the impact of the target signal on the statistical characteristics of the pixel time series intensity signal. μ and σ represent the mean and standard deviation of the original pixel time series intensity signal. and are the corrected mean and standard deviation.

[0088] The pixel time series intensity signal after filling for:

[0089]

[0090] δ represents the length of the padding signal, and Concate represents the signal concatenation operation.

[0091] Step A3: Perform Gaussian filtering on the pixel time series intensity signal to smooth the background noise.

[0092]

[0093] σ g represents the standard deviation of Gaussian filtering, X G (t) represents the pixel time series intensity signal after filtering.

[0094] Step A4: Standardize the filtered pixel time series intensity signal to make it satisfy the standard normal distribution.

[0095]

[0096] μ G and σ G Represents the mean and standard deviation of the pixel time series intensity signal after Gaussian filtering.

[0097] Step 2: Reconstruction of pixel time series intensity signal

[0098] Step B: Use time series analysis technology to process all input pixel time series intensity signals, suppress background and noise signals, and reconstruct the original pixel time series intensity signals into pixel time series intensity signals containing only target signals. The specific process includes:

[0099] Step B1: Initialize hyperparameters, including target signal width W, local standard deviation threshold θ σ , Standard deviation bimodal maximum difference D p wait.

[0100] Step B2: Perform the following steps for each pixel time series intensity signal input.

[0101] Step B3: Using the target signal width as a sliding window, calculate the local standard deviation within the window to form a local standard deviation signal.

[0102]

[0103] μW represents the mean value within the sliding window.

[0104] Step B4: Analyze the local standard deviation signal element by element. If an element is greater than the set standard deviation threshold, that is, S σ (k)>θ σ , then the abnormal standard deviation may be caused by the target signal or the noise signal. It is necessary to determine the statistical characteristics and shape characteristics of the local standard deviation signal after index k to determine whether it is the target signal. At this time, the processing flow goes to step B5. Otherwise, if the element is less than or equal to the set standard deviation threshold, that is, S σ (k)≤θ σ , then the area is the background signal, traverse the next element and go to step B4.

[0105] Step B5: Extreme point detection and peak analysis. If the abnormal standard deviation is generated by the target signal, then in the local standard deviation signal S σ (k) will form a double-peaked signal, such as Figure 1 The method for determining whether the local standard deviation signal after index k satisfies the bimodal feature is to determine whether the target signal exists. σ (k) whether there are two maximum points k1, k3 and a minimum point k2 in the local area, and the standard deviation difference between the two maximum points should be less than the maximum difference of the standard deviation peaks, that is, |S σ (k1)-S σ (k3)|<D p k1, k2, k3 are the element indices corresponding to the extreme points. If the above conditions are met, it can be considered that there is a target signal in the local area, and go to step B6. Otherwise, update the element index and go to step B4.

[0106] Step B6: Reconstruct the pixel time series intensity signal containing only the target signal. After the preset conditions in step B5 are met, it can be considered that the target signal exists in the local area of ​​the original pixel time series intensity signal. The start and end indexes of the target signal are the maximum point k, and the element index k4 corresponding to the foot of the second peak of the bimodal pattern in the local standard deviation signal. The elements between k and k4 in the original pixel time series intensity signal are retained. Then the element index is updated and go to step B4. The elements are iterated until the pixel time series intensity signal processing is completed. The reconstruction process of the pixel time series intensity signal is shown in formula (7).

[0107]

[0108] Step 3: Detection Image Reconstruction

[0109] Step C: After obtaining all reconstructed pixel time series intensity signals, extract the data at the specified frame of all pixel time series intensity signals, and reconstruct the detection image according to the coordinates when the pixel time series intensity signals were extracted in step A1. The specific process includes:

[0110] Step C1: extracting data at a specified frame of all pixel time series intensity signals frame by frame;

[0111] Step C2: reconstruct the data of all pixel time series intensity signals at the specified frame into a two-dimensional matrix I(k) according to the coordinates:

[0112]

[0113] Where M represents the number of rows of the image, N represents the number of columns of the image, and M*N represents the number of pixels in the image, which is also the number of pixel time series intensity signals; represents the first pixel time series intensity signal, represents the last pixel time series intensity signal; K represents the Kth frame; Reshape represents the reconstruction of a one-dimensional vector into a two-dimensional matrix;

[0114] Step C3: Normalize the two-dimensional matrix to between 0 and 255 to obtain a detection image.

[0115]

[0116] Step 4: Dense Object Segmentation

[0117] Step D: After obtaining the detection image, use the density clustering method to process each frame of the detection image to obtain all the pixels corresponding to each target. The specific process includes:

[0118] Step D1: Initialize the density clustering hyperparameters, which include the neighborhood radius ε, the minimum number of points P within the neighborhood radius M ;

[0119] Step D2: Starting from any pixel p with non-zero intensity value, mark it as “visited” and determine whether the neighborhood of p has at least P M If there is a pixel with non-zero intensity value, it is judged as a core pixel, and a new target cluster C is created for pixel p, and all pixels with non-zero intensity value in the neighborhood of p are put into the candidate target cluster G; otherwise, it is marked as a noise point;

[0120] Step D3: Iteratively add the non-zero intensity pixels in the candidate target cluster G that do not belong to other target clusters to the target cluster C. For the non-zero intensity pixel p' marked as "unvisited" in G, mark it as "visited", and check the neighborhood of the non-zero intensity pixel p'. If p' is a core pixel, add all the non-zero intensity pixels in the neighborhood of p' to the candidate target cluster G. Continue to add non-zero intensity pixels to the target cluster C until the target cluster C cannot be expanded, that is, the candidate target cluster G is empty, and the target cluster C is completely generated.

[0121] Step D4: Randomly select the next non-zero intensity pixel that has not been visited from the remaining non-zero intensity pixels, and go to step D3 until all non-zero intensity pixels are visited, and then end this step.

[0122] Step 5: Obtain the target bounding box

[0123] Step E: After obtaining all the pixels corresponding to each target, that is, the pixel set of the target, use the matrix externalization method to obtain the bounding box of the target, and add the bounding box to the original image to form the final detection result. The specific process includes:

[0124] Step E1: Calculate the target centroid based on all pixels corresponding to the target. The calculation formula is:

[0125]

[0126] Where (x i ,y i ) are the pixel coordinates, and n is the number of pixels in the target pixel set.

[0127] Step E2: Based on the target pixel set, construct the target covariance matrix. This matrix reflects the distribution characteristics of the target in different directions. The covariance matrix calculation formula is:

[0128]

[0129] is the centroid coordinate obtained in step E1

[0130] Step E3: Perform eigenvalue decomposition on the target covariance matrix to determine the principal component directions. These principal component directions will be used to determine the major and minor axes of the bounding box.

[0131] Step E4: Calculate the width and height of the bounding box based on the result of eigenvalue decomposition. This is usually achieved by selecting the two largest eigenvalues, which represent the distribution range of the target along the principal component direction.

[0132] Step E5: Using the centroid and principal component directions, calculate the coordinates of the four corner points of the bounding box. This can be done by combining the principal component vector with the centroid coordinates and determining the lengths of the major and minor axes based on the eigenvalues.

[0133] Step E6: Draw the calculated bounding box corner points onto the original image. Use the drawing function of the image processing library to draw a rectangular box according to the corner point coordinates.

[0134] Step E7: Repeat the above steps to generate bounding boxes for all detected objects and add them to the original image to form a complete detection result.

[0135] Example 2

[0136] Embodiment 2 of the present invention provides a dense dim small target detection system based on time series analysis, which is implemented based on the method of embodiment 1 and includes:

[0137] A preprocessing module is used to preprocess the received image sequence to obtain a pixel time-series intensity signal corresponding to each pixel;

[0138] A pixel time series intensity signal reconstruction module is used to process each pixel time series intensity signal using a time series analysis technique, and repeat this step until all reconstructed pixel time series intensity signals containing only the target signal are obtained;

[0139] A detection image reconstruction module is used to obtain signal data corresponding to each frame of all reconstructed signals according to the frame number, and reconstruct the detection image corresponding to the frame;

[0140] The dense target segmentation module is used to process the detection image using density clustering, cluster discrete target pixel points into targets, and eliminate false alarm points;

[0141] The target bounding box acquisition module is used to obtain all pixels corresponding to the target, obtain the target bounding box by using the matrix external bounding box method, and add the bounding box to the original image to obtain the detection result of the frame. The process is repeated until the detection results of all frames are obtained.

[0142] Summary:

[0143] The present invention provides a method for detecting dense and small targets based on time series analysis, taking the pixel time series intensity signal formed by a single pixel over time as the processing object, the dense and small targets in the field of view will generate multiple target signals on the pixel time series intensity signal, converting the dense and small target detection problem in the spatial domain into the pixel time series intensity signal reconstruction problem in the time domain, using the time series analysis technology to suppress the background signal and reconstruct the target signal, and inverting the detection result according to the reconstructed pixel time series intensity signal. The present invention can effectively solve the problem of dense and small target detection and achieve high-performance detection.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A dense dim target detection method based on time series analysis, comprising: Step A: preprocess the received image sequence to obtain the pixel time series intensity signal corresponding to each pixel; Step B: Using time series analysis technology, reconstruct each pixel time series intensity signal, and repeat this step until all reconstructed pixel time series intensity signals containing only the target signal are obtained; Step C: obtaining the signal data corresponding to each frame in the reconstructed pixel time series intensity signal containing only the target signal according to the frame number, and reconstructing the detection image corresponding to the frame; Step D: Use density clustering to process the detection image, cluster discrete target pixel points into targets, and eliminate false alarm points; Step E: After obtaining all the pixels corresponding to the target, the bounding box of the target is obtained by using the matrix external bounding box method, and the bounding box is added to the original image to obtain the detection result of the frame. This step is repeated until the detection results of all frames are obtained; The step B comprises: Step B1: Initialize hyperparameters, including target signal width W, local standard deviation threshold θ σ The maximum difference between the two peaks and the standard deviation D p ; Step B2: Perform the following steps for each pixel time series intensity signal input: Step B3: Using the target signal width W as the sliding window, calculate the local standard deviation within the window to form a local standard deviation signal S σ (k): Among them, μ W represents the mean value in the sliding window, k represents the element index of the local standard deviation signal, and t represents the pixel time series intensity signal X G (t) element index; Step B4: local standard deviation signal S σ (k) Perform element-by-element determination: If S σ (k) is greater than the set standard deviation threshold θ σ , go to step B5; otherwise, traverse the next element, that is, add 1 to k, until the element-by-element determination is completed; Step B5: Perform extreme point detection and peak analysis: If S σ There are two maximum points k1, k3 and one minimum point k2 in the local area of ​​(k), and |S σ (k1)-S σ (k3)|<D p , where k1, k2, k3 are the element indices corresponding to the extreme points, D p is the maximum difference between the two peaks of the standard deviation; if there is a target signal in the local area, go to step B6; otherwise, add 1 to k and go to step B4; Step B6: The start and end indexes of the target signal are the element index k4 corresponding to the maximum point k1 and the foot of the second peak of the bimodal pattern in the local standard deviation signal; the elements between k1 and k4 in the original pixel time series intensity signal are retained, thereby realizing the pixel time series intensity signal X G (t), and the reconstructed signal S T (t): Traverse the next element, that is, add 1 to k, and go to step B4; iterate the elements until the processing of the pixel time series intensity signal is completed.

2. The method for detecting dense and small targets based on time series analysis according to claim 1, characterized in that: The step A comprises: Step A1: Obtain the pixel time series intensity signal corresponding to each pixel according to the coordinates; Step A2: Use the corrected mean and standard deviation of the pixel time series intensity signal as the mean and standard deviation of the filling signal to generate the filling signal according to the normal distribution Among them, α and β are artificially set hyperparameters used to offset the impact of the target signal on the statistical characteristics of the pixel time series intensity signal; μ and σ represent the mean and standard deviation of the original pixel time series intensity signal. and is the corrected mean and standard deviation, N() indicates normal distribution; Get the pixel time series intensity signal after filling for: Among them, δ represents the length of the padding signal, and Concate represents the signal concatenation operation; Step A3: Gaussian filtering is performed on the pixel time series intensity signal. The filtered pixel time series intensity signal X G (t) is: Among them, σ g is the standard deviation of Gaussian filtering, which is used to represent the pixel time series intensity signal after filtering; Step A4: Filter the pixel time series intensity signal X G (t) Standardize it to satisfy the standard normal distribution: Among them, μ G and σ G Represents the mean and standard deviation of the pixel time series intensity signal after Gaussian filtering.

3. The method for detecting dense and small targets based on time series analysis according to claim 1, characterized in that: The step C comprises: Step C1: extracting the data at the specified frame of the reconstructed pixel time series intensity signal containing only the target signal frame by frame; Step C2: reconstruct the data at the specified frame of all reconstructed pixel time series intensity signals containing only the target signal into a two-dimensional matrix I(K) according to the coordinates: Where M represents the number of rows of the image, N represents the number of columns of the image, and M*N represents the number of pixels in the image, which is also the number of pixel time series intensity signals; represents the first pixel time series intensity signal, Represents the second pixel time series intensity signal, represents the last pixel time series intensity signal, K represents the Kth frame, and Reshape() represents the reconstruction of a one-dimensional vector into a two-dimensional matrix; Step C3: Normalize the two-dimensional matrix to between 0 and 255 to obtain the detection image: Among them, Max() and Min() represent the maximum value and minimum value respectively.

4. The method for detecting dense small targets based on time series analysis according to claim 1, characterized in that: The step D comprises: Step D1: Initialize the density clustering hyperparameters, which include the neighborhood radius ε, the minimum number of points P within the neighborhood radius M ; Step D2: Starting from any pixel p with non-zero intensity value, mark it as "visited" and determine whether the neighborhood of p has at least P M If there are non-zero intensity pixels, the judgment is yes, then p is the core pixel, a new target cluster C is created for p, and all non-zero intensity pixels in the neighborhood of p are placed in the candidate target cluster G; otherwise, the non-zero pixel p is marked as a noise point; Step D3: Iteratively add the non-zero intensity pixels in the candidate target cluster G that do not belong to other target clusters to the target cluster C. For the non-zero intensity pixel p' marked as "unvisited" in G, mark it as "visited", and check the neighborhood of the non-zero intensity pixel p'. If p' is a core pixel, add all the non-zero intensity pixels in the neighborhood of p' to the candidate target cluster G, and continue to add non-zero intensity pixels to the target cluster C until the target cluster C cannot be expanded, that is, the candidate target cluster G is empty, and the target cluster C is completely generated; Step D4: Randomly select the next non-zero intensity pixel that has not been visited from the remaining non-zero intensity pixels, and go to step D3 until all non-zero intensity pixels are visited, and then end this step.

5. The method for detecting dense and small targets based on time series analysis according to claim 1, characterized in that: The step E comprises: Step E1: Calculate the target centroid based on all pixels corresponding to the target Among them, (x i ,y i ) is the pixel coordinate, n is the number of pixels in the target pixel set; Step E2: Based on the target pixel set, construct the target covariance matrix Cov: Among them, i represents the sequence number of the i-th pixel in the target pixel set, and T represents transposition; Step E3: performing eigenvalue decomposition on the target covariance matrix to determine the principal component direction; Step E4: Calculate the width and height of the bounding box according to the result of eigenvalue decomposition; Step E5: Using the centroid and the principal component directions, calculate the coordinates of the four corner points of the bounding box; Step E6: Draw the calculated bounding box corner points onto the original image; Step E7: Go to step E1 until bounding boxes are generated for all detected objects and added to the original image to form a complete detection result.

6. A system based on the dense dim small target detection method based on time series analysis according to claim 1, characterized in that: include: A preprocessing module is used to preprocess the received image sequence to obtain a pixel time-series intensity signal corresponding to each pixel; A pixel time series intensity signal reconstruction module is used to reconstruct each pixel time series intensity signal using a time series analysis technique, and repeat this step until all reconstructed pixel time series intensity signals containing only the target signal are obtained; A detection image reconstruction module is used to obtain the signal data corresponding to each frame in the reconstructed pixel time series intensity signal containing only the target signal according to the frame number, and reconstruct the detection image corresponding to the frame; The dense target segmentation module is used to process the detection image using density clustering, cluster discrete target pixel points into targets, and eliminate false alarm points; and The target bounding box acquisition module is used to obtain all pixels corresponding to the target, obtain the target's bounding box using the matrix external bounding box method, and add the bounding box to the original image to obtain the detection result of the frame. The process is repeated until the detection results of all frames are obtained.

Images