Space object tracking method and system based on multi-feature fusion

Abstract

The application discloses a space target tracking method and system based on multi-feature fusion, and the method comprises the following steps: obtaining a space target motion model according to the orbital motion of the space target; performing space target feature extraction on a single frame star map to obtain high-order feature information; initializing the track of the space target by combining the space target motion model and a logic-based track starting method to obtain an initial suspected track; and performing constraint on the initial suspected track according to the high-order feature information to obtain an initial target track. The system comprises a model construction module, a feature extraction module, an initialization module and a constraint module. By using the application, the overall real-time performance, accuracy and robustness of the algorithm are improved, and the real track of the target is determined in a shorter frame. The application can be widely applied to the field of astronomical observation.

Description

Technical Field

[0001] This invention relates to the field of astronomical observation, and in particular to a method and system for tracking space targets based on multi-feature fusion. Background Technology

[0002] During celestial observation, rapid and accurate detection of space targets in optical images is a crucial issue that must be addressed. Traditional space target tracking and association algorithms often only consider the motion information of the target, achieving trajectory association through multi-frame matching, while ignoring the target's own distribution characteristics. In sky regions with dense star formations, severe occlusion, and sticking phenomena, traditional space target tracking and association algorithms suffer from numerous problems such as missed trajectory detection, false detection, and target occlusion during target tracking. These problems interfere with the algorithm's determination of the target's true trajectory. Therefore, traditional space target tracking and association algorithms often require more frames during target tracking to eliminate false alarms from stars and noise, filter out reliable moving targets, and thus determine the target's true trajectory. Summary of the Invention

[0003] To address the aforementioned technical problems, the objective of this invention is to provide a spatial target tracking method and system based on multi-feature fusion, which can improve the overall real-time performance, accuracy, and robustness of the algorithm through multi-feature fusion of the target, and achieve confirmation of the target's true trajectory in shorter frames.

[0004] The first technical solution adopted in this invention is a spatial target tracking method based on multi-feature fusion, comprising the following steps:

[0005] Based on the orbital motion of the space target, a motion model of the space target is obtained;

[0006] Spatial target features are extracted from a single frame of star map to obtain high-order feature information;

[0007] By combining the spatial target motion model with a logic-based trajectory initiation method, the spatial target points are initialized to obtain an initial suspected trajectory;

[0008] The initial suspected trajectory is constrained based on high-order feature information to obtain the initial target trajectory;

[0009] The trajectory preservation algorithm based on Kalman filtering predicts and tracks the initial target trajectory and maintains the extension of the initial target trajectory to obtain the target trajectory.

[0010] Furthermore, the step of obtaining the motion model of the space target based on its orbital motion specifically includes:

[0011] The orbital motion of a space target is projected onto a planar image to obtain the orbital projection of the space target;

[0012] The displacement and angular velocity characteristics of the space target projection are calculated to obtain the space target motion model;

[0013] Furthermore, the motion trajectory of the target displacement characteristic on the image can be regarded as a straight line, the motion of the target angular velocity characteristic in the image plane can be regarded as uniform motion in a short time, and the spatial motion model can be approximated as uniform linear motion.

[0014] Furthermore, the step of extracting spatial target features from a single-frame star map to obtain higher-order feature information specifically includes:

[0015] The single-frame star map is preprocessed to remove noise and stray light regions, resulting in a preprocessed star map.

[0016] Suspected targets are extracted from the preprocessed star map and local normalization is performed to obtain the suspected target region;

[0017] The suspected target region is used as a model input into the backbone network for forward propagation to obtain the target's feature vector;

[0018] The target feature vector is classified using a fully connected layer and a normalized exponential function to obtain target classification information;

[0019] Principal component analysis is used to reduce the dimensionality of the output feature vector and extract the most significant 3D features to obtain higher-order feature information.

[0020] This optimization step divides the image into target areas and interference areas such as noise and stars, achieving preliminary elimination of interference such as noise and stars. Further screening of suspected target areas based on target classification information helps to reduce the area of ​​subsequent initialization of spatial target points.

[0021] Furthermore, the step of initializing the spatial target point by combining the spatial target motion model with a logic-based trajectory initiation method to obtain an initial suspected trajectory specifically includes:

[0022] Based on the spatial motion model, targets are divided into first-class targets and second-class targets, and target points are divided into first-class target points and second-class target points;

[0023] Confirm the target point type of the target point in the previous frame. Based on the first type of target and the second type of target, use the target point in the current frame as the center and search for target points within the range of different search radii in the second frame image. If the target point is of the first type, skip the target point. If the target point is of the second type, set a tracking gate at the corresponding position.

[0024] Search for target points within the radius of the corresponding location tracking gate. If no second type target point exists, interrupt the search and matching. If a second type target point exists, take the nearest second type target point at the corresponding location as the matching point and generate an initial suspected trajectory.

[0025] This optimization process reduces the impact of whole-pixel deviation caused by coarse positioning, reduces computational load, improves detection efficiency, and avoids the situation where the same trajectory is repeatedly detected.

[0026] Furthermore, the step of introducing higher-order feature information to constrain the initial suspected trajectory to obtain the initial target trajectory specifically includes:

[0027] The initial confidence level is set based on higher-order feature information, and an initial value is given;

[0028] Based on the target's original motion state, predict the target's location and set up a tracking gate at the predicted location. If there is no target point within the radius of the tracking gate, add an empty point as the target point at the predicted location.

[0029] The initial suspected trajectory is used to search and match subsequent images, and the confidence level of the initial suspected trajectory is adjusted according to the search and matching information.

[0030] When the confidence level of the initial suspected trajectory is determined to be 0, the search and matching of the initial suspected trajectory is terminated.

[0031] When the confidence level of the initial suspected trajectory meets a certain threshold, the initial target trajectory is generated.

[0032] This optimization process reduces the impact on trajectory matching caused by targets being obscured by stellar targets or intersecting with other space targets and thus not being detected. It also eliminates trajectory misalignment caused by three random, unrelated points that happen to match a suspected trajectory, further reducing interference from stellar targets and noise.

[0033] Furthermore, the step of the trajectory preservation algorithm based on Kalman filtering predicting and tracking the initial target trajectory and maintaining its extension to obtain the target trajectory specifically includes:

[0034] The Kalman filter-based method continues to match the initial target trajectory, taking the target point information of the initial target trajectory as input, and outputting the predicted position of the target in the next frame of the image after the last frame;

[0035] A tracking gate is set up at the target prediction location to verify and confirm the target points in the prediction area, and the detection results are obtained.

[0036] Based on the detection results, the target point information of the next frame after the last frame is determined. If no target point is detected in the prediction area, an empty point is introduced at the prediction position, and the target point information of the next frame after the last frame is corrected to the prediction position information. If a target point is detected in the prediction area, the target point information of the next frame after the last frame is corrected to the position information of that target point.

[0037] Using the target point correction information of the next frame after the last frame as input, the Kalman filter method is repeated until all star charts are matched to generate the target trajectory.

[0038] This optimization process yields stable tracking results, enabling the continuous maintenance of the trajectory.

[0039] The second technical solution adopted in this invention is: a spatial target tracking system based on multi-feature fusion, comprising:

[0040] The model building module is used to convert the orbital motion of a space target into two dimensions, thereby obtaining a motion model of the space target.

[0041] The feature extraction module is used to extract spatial target features from a single frame of star map to obtain high-order feature information;

[0042] The initialization module is used to initialize the spatial target points in conjunction with the spatial target motion model to obtain the initial suspected trajectory;

[0043] The constraint module is used to constrain the initial suspected trajectory based on high-order feature information to obtain the initial target trajectory;

[0044] The beneficial effects of the method and system of this invention are as follows: This invention utilizes the motion model of the target on the star map, classifies the suspected target region in the single frame star map through a feature extraction network, and extracts the high-order feature information of the target; constrains the initial suspected trajectory of the target motion extracted based on the trajectory association algorithm through the high-order feature information of the target, and obtains the initial target trajectory, thereby improving the overall real-time performance, accuracy and robustness of the algorithm, and realizing the confirmation of the true trajectory of the target in shorter frames. Attached Figure Description

[0045] Figure 1 This is a flowchart of the steps of the spatial target tracking method based on multi-feature fusion of the present invention;

[0046] Figure 2 This is a structural block diagram of the spatial target tracking system based on multi-feature fusion of the present invention;

[0047] Figure 3 This invention provides a spatial target orbital motion diagram based on a multi-feature fusion spatial target tracking method.

[0048] Figure 4This is a space target orbit projection image based on the multi-feature fusion space target tracking method of the present invention;

[0049] Figure 5 This is the initial suspected trajectory matching map of the second frame, where the header of the first suspected trajectory is a non-matching point.

[0050] Figure 6 The first frame shows the suspected trajectory header as the suspected matching point, and the second frame shows the initial suspected trajectory matching map.

[0051] Figure 7 This is a comparison chart of the detection results of the spatial target tracking method based on multi-feature fusion and the supporting software of the acquisition device according to the present invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0053] like Figure 1 As shown, this invention provides a spatial target tracking method based on multi-feature fusion, which includes the following steps:

[0054] S1. Based on the orbital motion of the space target, obtain the motion model of the space target;

[0055] S1.1, such as Figure 3 As shown, since most space targets orbit in near-circular orbits and the Earth's rotational angular velocity has little effect on the target's motion, it is assumed that the target's orbit is circular, the Earth's rotational angular velocity is negligible, point O is the Earth's center, and target is the space target;

[0056] S1.2 Project the orbital motion of the space target onto the planar image to obtain the orbital projection of the space target, such as... Figure 4 As shown, considering the limiting case, when the image plane Ω is completely parallel to the orbital plane of the space target, the target trajectory crosses the entire camera's field of view. In this case, the target trajectory has a limiting deviation δh on the image plane, where O is the Earth's center, P is the camera's location, R is the distance from the Earth's center to the Earth's surface, and r is the distance from the space target to the Earth's center. Let θ be the field of view of the camera, θ be the angle through which the target's trajectory passes in the image plane, A, B, and C be the left endpoint, center, and right endpoint of the target's trajectory on the image plane, respectively, Ω be the imaginary plane parallel to the orbital plane, Ω' be the actual image plane, and α be the angle between the imaginary plane and the actual image plane.

[0057] S1.3 Calculate the displacement and angular velocity characteristics of the space target projection to obtain the space target motion model. Figure 4 It can be known that the distance r from the spatial target to the Earth's center is:

[0058]

[0059] so:

[0060]

[0061] Simultaneous existence:

[0062]

[0063]

[0064] Since the image plane cannot be parallel to the orbital plane, and considering that the maximum field of view of the camera is... have:

[0065]

[0066] Substituting the values ​​and solving the problem, we can see that, taking an image plane with a size of 4096×4096 pixels as an example, L AC There are 4096 pixels, while the corresponding deviation value δh only corresponds to 11 pixels. Such a deviation can be ignored, so the motion trajectory of the target on the image can be regarded as a straight line.

[0067] The velocity of the space target in the orbit is:

[0068]

[0069] Among them, the gravitational constant u is 3.98600436 × 10 5 km·s -2 R e Let be the Earth's equatorial radius, and H be the orbital altitude of the space target. Assuming the angle between the target's velocity and the plane's intersection line is γ, then the horizontal and vertical projections of the target onto the image plane are respectively:

[0070]

[0071]

[0072] The magnitude of the actual angular velocity ω of the spatial target on the image plane is:

[0073]

[0074] As can be seen from formula (9), when γ = 90° or 180°, the velocity of the target in the spatial image is vertically upward. At this time, the angular velocity of the target reaches its minimum value, and the angular acceleration reaches its maximum value.

[0075]

[0076]

[0077] When γ = 0° or 270°, the direction of the space target's velocity is parallel to the direction of the intersection of the two planes. At this time, the space target's angular velocity reaches its maximum value, and its angular acceleration reaches its minimum value.

[0078]

[0079] a min =0 (13)

[0080] As can be seen from equations (10), (11), (12), and (13) above, the angular acceleration of the spatial target is much smaller than its angular velocity. In a short time, the change in its angular velocity can be ignored. Therefore, the motion of the target in the image plane can be considered as uniform motion in a short time.

[0081] In summary, the motion of a spatial target within the image plane in a short period of time can be considered as uniform motion, and the trajectory is approximately a straight line. That is, the motion model of a spatial target within the image plane can be approximated as uniform linear motion.

[0082] S2. Extract spatial target features from a single-frame star map to obtain high-order feature information;

[0083] S2.1. Preprocess the single-frame star map by using a median filtering algorithm to remove noise from the original image and using a SEP algorithm to suppress stray light regions in the original image to obtain a preprocessed star map.

[0084] S2.2 Extract suspected targets from the preprocessed star map, extract local maxima and the r×r region around the local maxima, and perform local normalization to obtain the suspected target region;

[0085] S2.3. To model the parallel computing characteristics of suspected target regions, all suspected target regions in the figure are input as a batch into the backbone network for forward propagation to obtain the feature vector of the target.

[0086] S2.4 Calculate the target feature vector using a fully connected layer and a normalized exponential function, divide the suspected target region, and obtain classification information. If the classification information is 1, the region is the target region; otherwise, it is a noise, star, or other interference region.

[0087] S2.5. Principal component analysis is used to reduce the dimensionality of the output feature vector, extract the most significant 3D features, and obtain high-order feature information to characterize the imaging characteristics of the target. This information is output together with the classification information for further confirmation of the target and trajectory association.

[0088] S3. Initialize the spatial target points by combining the spatial target motion model with the logic-based trajectory initiation method to obtain the initial suspected trajectory;

[0089] S3.1 Considering the varying distances between targets, the velocities of different targets in the star map differ significantly. While high-speed targets can still be approximated as a straight line without ignoring target positioning errors, pixel deviations have a substantial impact on low-speed targets, making it difficult to approximate them as a straight line. Based on this, this method improves upon the logic-based trajectory initiation method by proposing a hierarchical matching trajectory initiation algorithm based on multi-feature fusion. The specific improvements are as follows:

[0090] S3.1-1. Considering the situation where the target in the star map is not detected during continuous movement, a "ghost point" mechanism is introduced;

[0091] Targets in a star map are constantly moving. During their movement, they may pass by stellar targets, be occluded by stellar targets, or intersect and be occluded by other space targets. For some extremely faint targets, they may also temporarily go undetected due to changes in lighting. However, this disappearance process is often very brief, generally lasting only one to three frames. Moreover, the target's motion state remains basically unchanged before and after disappearance. To address this, a "ghost point" mechanism is introduced. If a target disappears, its location is predicted based on its original motion state, a ghost point is added at that location, and the number of consecutive ghost points for that trajectory is recorded. If the trajectory re-tracks the target, the number of consecutive ghost points is set to 0. If the number of consecutive ghost points reaches a certain threshold, the trajectory is considered abnormal and is interrupted.

[0092] S3.1-2. Considering the different target speeds, a graded speed mechanism is introduced;

[0093] Considering the varying speeds of the targets, they are categorized into multiple speed levels, designated as low-speed, medium-speed, and high-speed targets. Low-speed targets are the first type of targets in the claims, and medium-speed targets are the second type. In the low-speed category, the target's speed ranges from 0 to 5 pixels. This category includes stellar targets and space targets with very low speeds. In a short time, due to the extremely low speed, it is difficult to distinguish targets from stars, requiring long-term displacement accumulation to determine the target type. In the medium-speed category, the target's speed ranges from 5 to 60 pixels. The movement of these targets differs significantly from noise and stellar targets, making them easy to distinguish. In the high-speed category, the target's speed exceeds 60 pixels. Due to the excessive speed, the target exhibits a long trail, making its appearance very distinct, appearing as a "line." The target's direction and speed can be determined based on the size of the trail, eliminating the need for trajectory initialization using a trajectory initiation method.

[0094] S3.1-3. In order to improve detection efficiency and avoid the same trajectory being detected repeatedly, a target point classification mechanism is introduced;

[0095] To improve detection efficiency and avoid the same trajectory being detected repeatedly, target points are divided into unmatched points, suspected matching points, and matched points. The matched points are the first type of matched points in the claims, and the unmatched points and suspected matching points are the second type of matched points in the claims. Unmatched points are target points that have not been included in a suspected trajectory, suspected matching points are target points that have been included in a suspected trajectory, and matched points are points that have been included in a trajectory that has been confirmed as a spatial target. Different initial suspected trajectory generation strategies are used for different points.

[0096] S3.2 Based on the spatial motion model, the target and target point are distinguished and classified according to the hierarchical matching trajectory initiation algorithm based on multi-feature fusion, and an initial suspected trajectory is generated;

[0097] S3.2-1. Based on different target types, different trajectory matching mechanisms are formulated. High-speed targets are directly matched with existing high-speed target trajectories. If the match is successful, the target is recorded into the existing trajectory. If the match fails, a new trajectory is generated directly without the need to generate an initial suspected trajectory. Medium-speed targets and low-speed targets are searched and matched separately. Due to the difference in speed, the two types of targets do not interfere with or affect each other.

[0098] S3.2-2. Based on the target point classification mechanism, perform correlation matching on the target points within the search range to generate an initial suspected trajectory;

[0099] like Figure 5As shown, when the target point in the first frame is an unmatched point, in the second frame image, with the corresponding position of the target point in the first frame as the center, the search radius is set with two speed ranges: low speed and medium speed. If there is a target within the search range, the category of the target point is judged. If the target point is a matched point, it is skipped. If the target point is an unmatched point or a suspected matched point, the position of the target in the third frame is predicted based on the position of the target point. A tracking gate is set at the predicted position. If there is a non-matched target within the radius of the tracking gate, the target point closest to the predicted position is taken as the matching point, and an initial suspected trajectory is generated. The information of the three target points is updated to the suspected matching point. If there is no non-matched target within the radius of the tracking gate, the search and matching of the trajectory is interrupted.

[0100] like Figure 6 As shown, when the target point in the first frame is a suspected matching point, in the second frame image, with the corresponding position of the target point in the first frame as the center, the search radius is set at both low and medium speeds. If a target exists within the search range, the target point's category is determined. If the target point is a matched point, it is skipped; if the target point is an unmatched point, the target position in the third frame is predicted based on the target point's location. A tracking gate is set at the predicted position. If there is a non-matched target within the tracking gate radius, the target point closest to the predicted position is taken as the matching point, generating an initial suspected trajectory. The information of the three target points is updated to the suspected matching point. If there is no non-matched target within the tracking gate radius, the search and matching of the trajectory is interrupted. If the target point is a suspected match point, it is necessary to backtrack the information of the frame before the first frame. Based on the positions of the target points in these two frames, the position of the target in the previous frame is predicted in reverse. A tracking gate is set at this position. If there is a suspected match point within the tracking gate, it is considered that the three target points have been included in the same initial suspected trajectory, and no repeated matching is performed. If there is no target within the tracking gate or the existing targets are all non-suspected match points, the search and matching continues. Based on the target point information, the position of the target in the third frame is predicted, and a tracking gate is set at the predicted position. If there is a non-matched target within the radius of the tracking gate, the target point closest to the predicted position is taken as the match point, an initial suspected trajectory is generated, and the information of the three target points is updated to the suspected match point.

[0101] S4. Constrain the initial suspected trajectory based on high-order feature information to obtain the target trajectory;

[0102] S4.1. Due to the large number of stellar targets, space targets, and noise points in the image, the occurrence of randomness is likely. Three unrelated points may happen to meet the conditions for generating a suspected trajectory. In this case, the generated trajectory needs further search and confirmation. Therefore, an initial confidence level δ0 is set based on high-order feature information. The expression for the initial confidence level δ0 is as follows:

[0103]

[0104]

[0105] in, D represents the high-dimensional feature vectors of the three target points in the initial trajectory; c Input feature vector The cosine distance represents the feature similarity between two targets; T c If the manually set similarity threshold satisfies the similarity condition of the feature vectors in the formula, then the initial confidence level of the corresponding trajectory is set.

[0106] S4.2 When the initial trajectory of the target satisfies the condition of formula (14), let the initial confidence of the trajectory be δ0. Based on the initial confidence, the initial suspected trajectory and subsequent images are tracked and matched. When the suspected trajectory successfully tracks a new target, the confidence of the suspected trajectory increases by 1 based on the initial confidence. When the subsequent tracking result of the suspected trajectory is an empty point, the confidence of the suspected trajectory decreases by 1. When the confidence satisfies the condition of 5, the suspected trajectory is considered credible and a new trajectory can be generated. When the confidence is zero, the search and matching of the suspected trajectory is terminated to obtain the target trajectory.

[0107] S5. The trajectory preservation algorithm based on Kalman filtering predicts and tracks the initial target trajectory and maintains the extension of the initial target trajectory to obtain the target trajectory.

[0108] S5.1 The Kalman filter-based method continues to match the initial target trajectory. When the target point information of the initial target trajectory is used as input, the predicted position of the target in the next frame of the last frame is output.

[0109] Kalman filtering is a filtering algorithm based on the Bayesian framework that satisfies certain assumptions. The recursive update equation for Bayesian filtering is:

[0110]

[0111] Where η is:

[0112]

[0113] use Replace the improper integral term in equation (16), that is:

[0114]

[0115] We can obtain:

[0116]

[0117] In the solution process, the first step is to solve... Let this be the prediction step, and then solve for... This is recorded as an update step.

[0118] During tracking, the following assumptions must be satisfied for updating and predicting using Kalman filtering:

[0119] The noise in the state transition matrix follows a Gaussian distribution:

[0120]

[0121] Right now:

[0122]

[0123] The noise in the observation matrix follows a Gaussian distribution:

[0124]

[0125] Right now:

[0126]

[0127] The initial state bel(x0) follows a Gaussian distribution:

[0128]

[0129] After satisfying the above assumptions, the prediction step and update step of the Bayesian algorithm are solved. The prediction step of the Bayesian algorithm is as follows:

[0130]

[0131] in

[0132] Therefore, we can conclude that:

[0133]

[0134]

[0135]

[0136] Kalman filter prediction estimates the target state vector X in the (k+1)th frame based on the first k observations Y. k+1 When the system is determined, assume T k For x k The estimated value The covariance matrix, T k 'For estimated value' With x k The error covariance matrix. The Kalman filter prediction process is as follows:

[0137] At time t0, initialization To find T0;

[0138] In t k At time t, the system's state prediction equation is:

[0139]

[0140] The system's state update equation is:

[0141]

[0142] Where K k The Kalman gain is calculated as follows:

[0143]

[0144]

[0145] T k = (1-K) k H k )T k (32)

[0146] Since the target can be considered to be moving at a constant velocity in a short period of time, according to the target's motion model, we can obtain:

[0147]

[0148] The state transition matrix of the target can be obtained as follows:

[0149]

[0150] Where Δt is the unit time interval, which can be considered as 1, the target observation matrix is:

[0151]

[0152] During target tracking, the process excitation noise W k and observation noise V k If it can be considered as Gaussian white noise with a mean of 0, then the process excitation noise W k The covariance matrix is:

[0153]

[0154] The covariance matrix of the observation noise is:

[0155]

[0156] S5.2 Predict the centroid position of the target in the next frame using the time update equation and state update equation in the Kalman filter.

[0157] S5.3 Set up a tracking gate at the target prediction location, verify and confirm the target points in the prediction area, and obtain the detection results;

[0158] S5.4. Based on the detection results, determine the target point information of the next frame after the last frame. If no target point is detected in the prediction area, introduce an empty point in the prediction position and correct the target point information of the next frame after the last frame to the prediction position information. If a target point is detected in the prediction area, correct the target point information of the next frame after the last frame to the position information of that target point.

[0159] S5.5. Using the target point correction information of the next frame after the last frame as input, perform a loop Kalman filter until all star charts are matched to generate the target trajectory.

[0160] like Figure 2 As shown, a spatial target tracking system based on multi-feature fusion includes:

[0161] The model building module is used to convert the orbital motion of a space target into two dimensions, thereby obtaining a motion model of the space target.

[0162] The feature extraction module is used to extract spatial target features from a single frame of star map to obtain high-order feature information;

[0163] The initialization module is used to initialize the spatial target points in conjunction with the spatial target motion model to obtain the initial suspected trajectory;

[0164] The constraint module is used to constrain the initial suspected trajectory based on high-order feature information to obtain the initial target trajectory.

[0165] To verify the effectiveness of the algorithm, the experiment used multiple sets of collected astronomical data for comparative verification. Specifically, in fixed tracking mode, data from different sky regions were collected to verify the robustness of the detection algorithm and trajectory association algorithm under different exposure times.

[0166] The data acquisition devices are shown in the table below:

[0167] Table 1 Data Acquisition Camera Parameters

[0168]

[0169]

[0170] The algorithm of this invention is used for multi-target tracking of detection data. For an image with a size of 4096×4096 pixels, the processing speed per frame is approximately 0.4 seconds, which can well meet the real-time requirements of engineering projects. The tracking results are compared with the processing results of the software accompanying the acquisition device, such as... Figure 7As shown, image data acquired from four different sky regions were processed. The second column shows the processing results of the acquisition equipment's software for images acquired with a 500ms exposure time. The last two columns show the processing results of the algorithm presented in this paper for images acquired with 100ms and 500ms exposure times, respectively. It can be seen from the images that:

[0171] 1) The algorithm in this paper has basically the same detection results for images with 100ms and 500ms exposure, and has good adaptability to changes in exposure time, with no false detections or missed detections.

[0172] 2) By comparing the processing results of the accompanying software for 500ms exposure images, the detection results of the algorithm in this paper include not only all the trajectories detected by the accompanying software, but also some additional trajectories, which are marked with blue dashed boxes in the image. Observation and verification show that the additional trajectories are all real targets, indicating that the algorithm in this paper has a stronger ability to detect and track targets in weak spaces and has a higher tracking success rate.

[0173] 3) When the target is relatively dark, such as in sky areas 3 and 4, the overall brightness in the image is low, and the accompanying software has a large number of broken tracks in the detection results. However, the algorithm in this paper can still maintain good tracking stability for dark targets, and there are fewer broken tracks.

[0174] To better describe the tracking performance of the algorithm presented in this chapter, it is compared with the Probability Hypothesis Density Filtering (CPHD) algorithm and the Generalized Label Multi-Bernoulli (GLMB) algorithm. The tracking performance is evaluated using three metrics: false detection rate (O), false miss rate (N), and number of interrupted trajectories (PM). Here, the false detection rate is the ratio of falsely detected trajectories to the total number of trajectories; the false miss rate is the ratio of falsely detected trajectories to the total number of trajectories; and the number of interrupted trajectories represents the number of trajectories interrupted due to tracking gaps or matching errors. The comparison results are shown in the table below:

[0175] Table 2 Comparison of Algorithm Tracking Performance

[0176]

[0177] As shown in Table 2, the CPHD algorithm's trajectory false detection rate, trajectory missed detection rate, and number of trajectory interruptions are all in the middle of the three algorithms. The GLMB algorithm has the lowest trajectory false detection rate, with no falsely detected trajectories, but its trajectory missed detection rate and number of interrupted trajectories are higher than the other two algorithms, indicating lower tracking success rate and tracking stability. The algorithm presented in this paper has no missed trajectories, indicating a higher tracking success rate; the number of interrupted trajectories is much lower than the other two algorithms, indicating better tracking stability; while maintaining a high tracking success rate and tracking stability, the algorithm in this paper also maintains a low trajectory false detection rate, indicating that the overall tracking performance of the algorithm presented in this paper is better than that of the CPHD and GLMB algorithms.

[0178] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0179] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A spatial target tracking method based on multi-feature fusion, characterized in that, Includes the following steps: Based on the orbital motion of the space target, a motion model of the space target is obtained; Spatial target features are extracted from a single frame of star map to obtain high-order feature information; By combining a spatial target motion model with a logic-based trajectory initiation method, spatial target points are initialized, and a graded velocity mechanism and a target point classification mechanism are introduced to obtain an initial suspected trajectory. The initial suspected trajectory is constrained based on high-order feature information to obtain the initial target trajectory; The step of extracting spatial target features from a single-frame star map to obtain high-order feature information specifically includes: The single-frame star map is preprocessed to remove noise and stray light regions, resulting in a preprocessed star map. Suspected target regions are obtained by extracting suspected targets from the preprocessed star map; The suspected target region is used as a model input into the backbone network for forward propagation to obtain the target's feature vector; The target's feature vector is classified using a fully connected layer and a normalized exponential function to obtain target classification information; Principal component analysis is used to reduce the dimensionality of the target's feature vectors and extract the most significant 3D features to obtain higher-order feature information. The step of constraining the initial suspected trajectory based on high-order feature information to obtain the initial target trajectory specifically includes: The initial confidence level is set based on higher-order feature information, and an initial value is given; Based on the target's original motion state, predict the target's location and set up a tracking gate at the predicted location. If there is no target point within the radius of the tracking gate, add an empty point as the target point at the predicted location. The initial suspected trajectory is used to search and match subsequent images, and the confidence level of the initial suspected trajectory is adjusted according to the search and matching information. If the confidence level of the initial suspected trajectory is determined to be 0, the search and matching of the initial suspected trajectory is terminated. When the confidence level of the initial suspected trajectory meets a certain threshold, the initial target trajectory is generated.

2. The spatial target tracking method based on multi-feature fusion according to claim 1, characterized in that, Also includes: The trajectory preservation algorithm based on Kalman filtering predicts and tracks the initial target trajectory and maintains the extension of the initial target trajectory to obtain the target trajectory.

3. The spatial target tracking method based on multi-feature fusion according to claim 1, characterized in that, The step of obtaining the motion model of the space target based on its orbital motion specifically includes: The orbital motion of a space target is projected onto a planar image to obtain the orbital projection of the space target; The displacement and angular velocity characteristics of the space target's orbital projection are calculated to obtain the space target's motion model.

4. The spatial target tracking method based on multi-feature fusion according to claim 1, characterized in that, The step of initializing the spatial target point by combining the spatial target motion model with a logic-based trajectory initiation method to obtain an initial suspected trajectory specifically includes: Considering the differences in distance and speed between targets, the logic-based trajectory initiation method is improved to obtain a hierarchical matching trajectory initiation algorithm based on multi-feature fusion; Based on the spatial motion model, the target and target point are distinguished and classified according to the hierarchical matching trajectory initiation algorithm based on multi-feature fusion, and an initial suspected trajectory is generated.

5. The spatial target tracking method based on multi-feature fusion according to claim 4, characterized in that, The step of generating an initial suspected trajectory based on a spatial motion model and a hierarchical matching trajectory initiation algorithm based on multi-feature fusion to differentiate and classify the target and target points specifically includes: Based on the spatial motion model, targets are divided into first-class targets and second-class targets, and target points are divided into first-class target points and second-class target points; Confirm the target point type of the target point in the previous frame. Based on the first type of target and the second type of target, use the target point in the current frame as the center and search for target points within the range of different search radii in the second frame image. If the target point is of the first type, skip the target point. If the target point is of the second type, set a tracking gate at the corresponding position. Search for target points within the radius of the corresponding location tracking gate. If no second type target point exists, interrupt the search and matching. If a second type target point exists, take the nearest second type target point at the corresponding location as the matching point and generate an initial suspected trajectory.

6. The spatial target tracking method based on multi-feature fusion according to claim 1, characterized in that, The initial confidence expression is as follows: in, D represents the high-dimensional feature vectors of the three target points in the initial trajectory; c Input feature vector The cosine distance represents the feature similarity between two targets; T c If the manually set similarity threshold satisfies the similarity condition of the feature vectors in the formula, then the initial confidence level of the corresponding trajectory is set.

7. The spatial target tracking method based on multi-feature fusion according to claim 2, characterized in that, The trajectory preservation algorithm based on Kalman filtering predicts and tracks the initial target trajectory, and maintains the extension of the initial target trajectory to obtain the target trajectory. This step specifically includes: The Kalman filter-based method continues to match the initial target trajectory, taking the target point information of the initial target trajectory as input, and outputting the predicted position of the target in the next frame of the image based on the initial target trajectory; A tracking gate is set up at the target prediction location to verify and confirm the target points in the prediction area, and the detection results are obtained. Based on the detection results, the target point information of the next frame of the initial target trajectory is determined, and the correction information is obtained; The target point correction information of the next frame of the initial target trajectory is used as input, and the Kalman filter method is used in a loop until all star charts are matched to generate the target trajectory.

8. A spatial target tracking system based on multi-feature fusion, characterized in that, For performing the spatial target tracking method based on multi-feature fusion as described in claim 1, including: The model building module is used to convert the orbital motion of a space target into two dimensions, thereby obtaining a motion model of the space target. The feature extraction module is used to extract spatial target features from a single frame of star map to obtain high-order feature information; The initialization module is used to initialize the spatial target points in conjunction with the spatial target motion model to obtain the initial suspected trajectory; The constraint module is used to constrain the initial suspected trajectory based on high-order feature information to obtain the initial target trajectory.

Images