A spatial high-maneuvering target tracking method based on deep correlation

Abstract

The present invention relates to a spatial high-maneuverable target tracking method based on deep association, including: designing a target detector based on a kernel function based on a two-dimensional high-resolution image sequence obtained by shooting the detection area; using the kernel function to reflect the similarity characteristics of data in high-dimensional space, the time domain data deep association algorithm distinguishes the signal of the moving target from the background clutter and noise signal in the time domain, and maps the time series corresponding to a certain pixel to the high-dimensional space to achieve continuous tracking of the spatial high-maneuverable target. This patent proposes a target tracking method based on an image sequence, which is conducive to the deep association method to mine the differences in the time series characteristics of the target and background clutter and noise, realize the detection and tracking of the target under low signal-to-noise ratio conditions, and solve the track association problem of high-maneuverable targets.

Description

Technical Field

[0001] The present invention relates to the technical field of space target tracking and detection, and in particular to a space high-maneuverability target tracking method based on depth association. Background Art

[0002] With the increasing level of modern aerospace science and technology, new high-maneuverable spacecraft have become a new development direction for the space power of aerospace powers. This type of target has complex movement patterns and strong high-speed maneuverability, which brings huge challenges to aerospace measurement and control and target perception: on the one hand, it is difficult for the existing technical system to include detection and stable tracking of high-speed maneuverable targets, and it is impossible to measure the spatial position and attitude of the target. On the other hand, the existing equipment does not have fast imaging capabilities and cannot measure the characteristics of high-maneuverable targets. The working mode of traditional radar equipment is mostly based on time sampling of continuous pulse accumulation, while high-speed maneuverable targets are sensitive to time and have the characteristics of large dynamics and fast fluctuations, which makes it difficult for the existing system to quickly obtain target measurement information. In addition, the detection signal of high-speed maneuverable targets in space generally has a low signal-to-noise ratio, and the target is easily submerged by background clutter. Even after noise suppression, it is difficult to detect and track high-maneuverable targets. Summary of the invention

[0003] In view of this, the present invention provides a spatial high-maneuvering target tracking method based on deep association, which can realize the detection and tracking of targets under low signal-to-noise ratio conditions and solve the track association problem of high-maneuvering targets.

[0004] In order to solve the above technical problems, the technical solution of the present invention includes the following contents:

[0005] S1. For the same detection area, N time series images are obtained along the time dimension.

[0006] S2. Perform normalization preprocessing on the time series images.

[0007] S3. Design a kernel function-based target deep correlation detector.

[0008] S4. According to the possible motion state of the target, select an appropriate sliding window size of the target depth association detector.

[0009] S5. Perform depth association on the pixels to be detected and the background pixels along the time series respectively.

[0010] S6. Statistically calculate the detection threshold based on the background pixel depth correlation result.

[0011] S7. Using the detection threshold to perform threshold judgment on the pixel points to be detected, so as to achieve continuous tracking of the high-mobility target in space.

[0012] Furthermore, in S1, for the same detection area, N time-series images are obtained along the time dimension, which specifically includes the following steps:

[0013] S1-1. According to the type of target to be detected, set a reasonable image resolution ρ and the accumulation time T corresponding to each frame of image f .

[0014] S1-2. Long accumulation time T for acquisition l The echo data is accumulated according to the single frame image time T f Divide and obtain N echo sequences, N = T l / T f .

[0015] S1-3, perform imaging processing on N echo sequences respectively to obtain N image sequences about the same detection area, where the two-dimensional length and width of each image are N a ,N r .

[0016] Further, S2, normalization preprocessing is performed on the time series images, comprising the following steps:

[0017] S2-1. Count the mean δ and standard deviation σ of the time series image.

[0018] S2-2, normalize the N images respectively. Before normalization, the pixel value of the i-th and j-th points of the n-th image is X n (i,j), then the normalized result is

[0019] Further, S3, design a target deep correlation detector based on kernel function, specifically:

[0020] The designed kernel function-based target depth correlation detector is f F (X,Y):

[0021]

[0022] where f F (X, Y) is the depth correlation result calculated for signal vectors X and Y, m is the sliding window length of the target depth correlation detector, x k is the kth signal in the signal vector X, y k is the kth signal in the signal vector Y.

[0023] Further, S4, according to the possible motion state of the target, a suitable sliding window size of the association detector is selected, specifically: the sliding window size of the designed target depth association detector is W=N / 4.

[0024] Further, S5, performing depth association on the pixels to be detected and the background pixels along the time series respectively, comprises the following steps:

[0025] S5-1. Assume that the pixel to be analyzed is the (i, j)th pixel, i is along the row direction, j is along the column direction, obtain the N pixel values ​​of the (i, j)th pixel point along the time series, and record them as D n ,n=1,2,...,N。

[0026] S5-2. Obtain all W signals of the signal vector X described by the target depth correlation detector according to the sliding window size, that is, X = D1, D2, ..., D W and all W signals in the signal vector Y, that is, Y = D2, D3, ..., D W+1 .

[0027] S5-3. According to the target depth correlation detector, calculate the depth correlation result of the two signal vectors X and Y in S5-2, denoted as K ij (1).

[0028] S5-4. Move the sliding window position with a step size of 1 to obtain all W signals in the signal vector X.

[0029] X=D2,D3,...,D W+1 and all W vectors in vector Y = D3, D4, ..., D W+2 , still according to the target depth association detector, calculate the depth association result, denoted as K ij (2).

[0030] S5-5. Move the sliding window position in step 1, repeat the depth correlation calculation of S5-4, and obtain the depth correlation result K ij (3),K ij (4),...,K ij (N-2W), a total of N-2W sliding windows need to be moved.

[0031] S5-6. Traverse all pixels in the image and repeat steps S5-1 to S5-5 to obtain the depth association result of each pixel.

[0032] Further, S6, statistically calculating a detection threshold according to the background pixel depth correlation result, including: setting the detection threshold to Thre=8.

[0033] Furthermore, in step S7, the threshold value judgment of the pixel to be detected is to normalize the depth correlation result of the pixel to be detected and compare it with the threshold value Thre=8.

[0034] Beneficial effects:

[0035] The spatial high-maneuverable target tracking method based on deep association of the present invention proposes a target tracking method based on image sequence, which is beneficial to the deep association method to mine the differences in time and space sequence characteristics between the target and the background, realize the detection and tracking of the target under low signal-to-noise ratio conditions, and solve the problem of track association of high-maneuverable targets.

[0036] That is, the kernel function can reflect the similarity of data in high-dimensional space. The time domain data deep association algorithm distinguishes the signal of the moving target from the background clutter signal and noise in the time domain, and maps the time series corresponding to a certain pixel to the high-dimensional feature space, thereby realizing continuous tracking of high-speed moving targets in space.

[0037] Embodiment 1

[0038] like Figure 1 As shown, a spatial high-maneuverability target tracking method based on deep association provided by an embodiment of the present invention includes the following steps:

[0039] S1, obtaining 100 low-resolution sequence images of the detection area along the time dimension;

[0040] The inverse SAR measured data of a certain radar equipment was used for preliminary verification. 100 inverse SAR images with a resolution of 100m were obtained in the detection area. The number of pixels in each frame was Na=100, Nr=100, and the target's moving speed was about 200m / s.

[0041] S2, normalizing and preprocessing the time series images;

[0042] Preferably, performing normalization preprocessing on the acquired inverse SAR image sequence includes:

[0043] S2-1, calculate the mean δ and variance σ of each image sequence;

[0044] S2-2, normalize the N images respectively. Before normalization, the pixel value of the i-th and j-th points of the n-th image is X n (i,j), then the normalized result becomes

[0045] S3. Design a kernel function-based target deep correlation detector;

[0046] Considering the target characteristics, the target depth correlation detector involved is

[0047] Where m is the sliding window length to be calculated in claim 4, x k is the kth signal in the signal vector X, y k is the kth signal in the signal vector Y.

[0048] S4. Select an appropriate sliding window size of the associated detector according to the possible motion state of the target.

[0049] The sliding window size of the correlation detector is set to W=N / 4=25.

[0050] S5, performing depth correlation on the pixels to be detected and the background pixels along the time series respectively;

[0051] The background pixels and the target pixels to be detected are deeply associated respectively. The specific association method is as follows:

[0052] S5-1. Assume that the pixel to be analyzed is the (i, j)th pixel (i is along the row direction, j is along the column direction), obtain the N pixel values ​​of the pixel along the time series, which are respectively recorded as D n ,n=1,2,...,N;

[0053] S5-2. Obtain all W signals of the signal vector X described in claim 4 according to the sliding window size, that is, X = D1, D2, ..., D W and all W signals in the signal vector Y, that is, Y = D2, D3, ..., D W+1 ;

[0054] S5-3. According to the target detector designed in claim 3, calculate the depth correlation result of the two signal vectors X and Y in S5-2, denoted as K ij (1);

[0055] S5-4. Move the sliding window position with a step size of 1 to obtain all W signals in the signal vector X

[0056] X=D2,D3,...,D W+1 and all W vectors in vector Y = D3, D4, ..., D W+2 , still according to the target detector designed in claim 3, calculate the depth association result, denoted as K ij (2);

[0057] S5-5. Move the sliding window position in step size 1, repeat the depth correlation calculation of S5-4, and obtain the correlation result K ij (3),K ij (4),...,K ij (N-2W), a total of N-2W sliding windows need to be moved;

[0058] S5-6. Traverse all pixels in the image, repeat steps S5-1 to S5-5, and obtain the depth association result of each pixel;

[0059] S6. Statistically calculate the detection threshold based on the background pixel depth correlation result;

[0060] Select 50*50 pixels as background pixels and calculate the average and standard deviation of the depth correlation results of these 2500 pixels. The specific steps include the following three steps:

[0061] S6-1. Calculate the average value α and variance β of the depth correlation results of all pixels in the background area;

[0062] S6-2. Normalize the depth correlation result of each pixel to

[0063] S6-3. Set the detection threshold to Thre=8;

[0064] S7. Perform threshold judgment on the pixels to be detected to achieve continuous tracking of highly maneuverable targets in space.

[0065] The depth correlation result of the pixel to be tested is normalized and compared with the threshold Thre=8.

[0066] Beneficial effects:

[0067] The spatial high-maneuverable target tracking method based on deep association of the present invention proposes a target tracking method based on image sequences, which is conducive to the deep association method to mine the differences in the temporal and spatial sequence characteristics of the target and the background, realize the detection and tracking of the target under the condition of low signal-to-noise ratio, and solve the track association problem of high-maneuverable targets. That is, the kernel function can reflect the similarity of data in high-dimensional space. The deep association algorithm of time domain data distinguishes the signal of the moving target from the background clutter signal and noise in the time domain, and maps the time series corresponding to a certain pixel to the high-dimensional feature space, thereby realizing the continuous tracking of the spatial high-speed moving target. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The flowchart schematically shows the method for tracking a spatial high-speed maneuvering target based on deep association of the present invention. DETAILED DESCRIPTION

[0069] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0070] Embodiment 1

[0071] like Figure 1 As shown, a spatial high-maneuverability target tracking method based on deep association provided by an embodiment of the present invention includes the following steps:

[0072] S1, obtaining 100 low-resolution sequence images of the detection area along the time dimension;

[0073] The inverse SAR measured data of a certain radar equipment was used for preliminary verification. 100 inverse SAR images with a resolution of 100m were obtained in the detection area. The number of pixels in each frame was Na=100, Nr=100, and the target's moving speed was about 200m / s.

[0074] S2, normalizing and preprocessing the time series images;

[0075] Preferably, performing normalization preprocessing on the acquired inverse SAR image sequence includes:

[0076] S2-1, calculate the mean δ and variance σ of each image sequence;

[0077] S2-2, normalize the N images respectively. Before normalization, the pixel value of the i-th and j-th points of the n-th image is X n (i,j), then the normalized result becomes

[0078] S3. Design a kernel function-based target deep correlation detector;

[0079] Considering the target characteristics, the target depth correlation detector involved is

[0080] Where m is the sliding window length to be calculated in claim 4, x k is the kth signal in the signal vector X, y k is the kth signal in the signal vector Y.

[0081] S4. Select an appropriate sliding window size of the associated detector according to the possible motion state of the target.

[0082] The sliding window size of the correlation detector is set to W=N / 4=25.

[0083] S5, performing depth correlation on the pixels to be detected and the background pixels along the time series respectively;

[0084] The background pixels and the target pixels to be detected are deeply associated respectively. The specific association method is as follows:

[0085] S5-1. Assume that the pixel to be analyzed is the (i, j)th pixel (i is along the row direction, j is along the column direction), obtain the N pixel values ​​of the pixel along the time series, which are respectively recorded as D n ,n=1,2,...,N;

[0086] S5-2. Obtain all W signals of the signal vector X described in claim 4 according to the sliding window size, that is, X = D1, D2, ..., D Wand all W signals in the signal vector Y, that is, Y = D2, D3, ..., D W+1 ;

[0087] S5-3. According to the target detector designed in claim 3, calculate the depth correlation result of the two signal vectors X and Y in S5-2, denoted as K ij (1);

[0088] S5-4. Move the sliding window position with a step size of 1 to obtain all W signals in the signal vector X

[0089] X=D2,D3,...,D W+1 and all W vectors in vector Y = D3, D4, ..., D W+2 , still according to the target detector designed in claim 3, calculate the depth association result, denoted as K ij (2);

[0090] S5-5. Move the sliding window position in step size 1, repeat the depth correlation calculation of S5-4, and obtain the correlation result K ij (3),K ij (4),...,K ij (N-2W), a total of N-2W sliding windows need to be moved;

[0091] S5-6. Traverse all pixels in the image, repeat steps S5-1 to S5-5, and obtain the depth association result of each pixel;

[0092] S6. Statistically calculate the detection threshold based on the background pixel depth correlation result;

[0093] Select 50*50 pixels as background pixels and calculate the average and standard deviation of the depth correlation results of these 2500 pixels. The specific steps include the following three steps:

[0094] S6-1. Calculate the average value α and variance β of the depth correlation results of all pixels in the background area;

[0095] S6-2. Normalize the depth correlation result of each pixel to

[0096] S6-3. Set the detection threshold to Thre=8;

[0097] S7. Perform threshold judgment on the pixels to be detected to achieve continuous tracking of highly maneuverable targets in space.

[0098] The depth correlation result of the pixel to be tested is normalized and compared with the threshold Thre=8.

[0099] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A spatial high-mobility target tracking method based on deep association, characterized in that: The steps include: S1. For the same detection area, N time series images are obtained along the time dimension; S2, performing normalization preprocessing on the time series images; S3. Design a kernel function-based target deep correlation detector; S4, according to the possible motion state of the target, select an appropriate sliding window size of the target depth correlation detector; S5, performing depth correlation on the pixels to be detected and the background pixels along the time series respectively; S6. Statistically calculate the detection threshold based on the background pixel depth correlation result; S7, using the detection threshold to perform threshold judgment on the pixel points to be detected, so as to achieve continuous tracking of the high-mobility target in space; In S1, for the same detection area, N time-series images are acquired along the time dimension, which specifically includes the following steps: S1-1. According to the type of target to be detected, set a reasonable image resolution ρ and the accumulation time T corresponding to each frame of image f ; S1-2. Long accumulation time T for acquisition l The echo data is accumulated according to the single frame image time T f Divide and obtain N echo sequences, N = T l / T f ; S1-3, performing imaging processing on N echo sequences respectively, obtaining N image sequences about the same detection area, wherein the two-dimensional length and width of each image are Na and Nr respectively; S2, performing normalization preprocessing on the time series images, comprises the following steps: S2-1, statistics of mean δ and standard deviation σ along the time series image; S2-2, normalize the N images respectively. Before normalization, the pixel value of the i-th and j-th points of the n-th image is X n (i,j), then the normalized result is S3, designing a target deep correlation detector based on a kernel function, specifically: The designed kernel function-based target depth correlation detector is f F (X,Y): where f F (X, Y) is the depth correlation result calculated for signal vectors X and Y, m is the sliding window length of the target depth correlation detector, x k is the kth signal in the signal vector X, y k is the kth signal in the signal vector Y.

2. The method according to claim 1, characterized in that: The S4, according to the possible motion state of the target, selects a suitable sliding window size of the association detector, specifically: the sliding window size of the designed target depth association detector is W=N / 4.

3. The method according to claim 2, characterized in that S5, performing depth correlation on the pixels to be detected and the background pixels along the time series respectively, comprises the following steps: S5-1. Assume that the pixel to be analyzed is the (i, j)th pixel, i is along the row direction, j is along the column direction, obtain the N pixel values ​​of the (i, j)th pixel point along the time series, and record them as D n ,n=1,2,...,N; S5-2. Obtain all W signals of the signal vector X described by the target depth correlation detector according to the sliding window size, that is, X = D1, D2, ..., D W and all W signals in the signal vector Y, that is, Y = D2, D3, ..., D W+1 ; S5-3. According to the target depth correlation detector, calculate the depth correlation result of the two signal vectors X and Y in S5-2, denoted as Kij(1); S5-4. Move the sliding window position with a step size of 1 to obtain all W signals X=D2, D3, ..., D in the signal vector X W+1 and all W vectors in vector Y = D3, D4, ..., D W+2 , still according to the target depth association detector, calculate the depth association result, denoted as K ij (2); S5-5. Move the sliding window position in step 1, repeat the depth correlation calculation of S5-4, and obtain the depth correlation result K ij (3),K ij (4),...,K ij (N-2W), a total of N-2W sliding windows need to be moved; S5-6. Traverse all pixels in the image and repeat steps S5-1 to S5-5 to obtain the depth association result of each pixel.

4. The method according to claim 3, characterized in that The step S6, statistically calculating the detection threshold according to the background pixel depth correlation result, includes: setting the detection threshold to Thre=8.

Images