A space non-cooperative target real-time tracking method and system thereof

Abstract

The application discloses a space non-cooperative target real-time tracking method and system, and belongs to the technical field of aerospace.The application comprises a lightweight space non-cooperative target detection module, a position extraction module, a detection frame matching module, a Kalman filtering module and a storage module.The application uses lightweight space non-cooperative target detection, utilizes MobileNetV3 as the main stem of an algorithm, simultaneously utilizes deep separable convolution to replace the traditional convolution layer of a PANet network, then reduces the parameter quantity of the model and improves the detection speed of the model through multiple upsampling, downsampling and feature fusion operations.The application obtains an initial target frame of non-Gaussian distribution according to an allowable domain, and improves the robustness and convergence of target tracking.The application uses an allowable domain method to determine the envelope where the space target position may exist according to one observation of current space non-cooperative target image information, and improves the accuracy and robustness of the selection of the frame area where the target is located through a mixed Gaussian distribution.

Description

Technical Field

[0001] This invention belongs to the field of aerospace technology and relates to a method and system for real-time tracking of non-cooperative targets in space. Background Technology

[0002] With the continuous development of space technology, countries around the world have carried out numerous missions, including communications, atmospheric and Earth resource observation. However, most spacecraft have become defunct and may collide with operational spacecraft, generating a large amount of space debris, which will threaten normally functioning spacecraft. In response to the increasingly severe space security situation and the scarcity of space resources, tasks such as on-orbit repair to extend the service life of defunct and malfunctioning spacecraft, and space debris cleanup have been proposed. Therefore, real-time tracking of non-cooperative targets in space is necessary, and obtaining information on these targets is crucial for subsequent missions.

[0003] Traditional target tracking methods utilize the edge or contour information of spatially non-cooperative targets to detect them and use this information as a template to match subsequent images for tracking. However, tracking failures are common due to variations in the scale of spatially non-cooperative targets, as well as interference from lighting and shadows, resulting in poor robustness. In recent years, the development of deep learning has led to improved accuracy and speed in target detection and tracking methods based on convolutional neural networks. However, existing methods are still relatively slow and lack robustness in occluded scenes and for tracking small targets. Therefore, it is necessary to propose a spatially non-cooperative target tracking method that is both powerful and fast, ensuring both accuracy and real-time performance. Summary of the Invention

[0004] To address the limitations of existing methods in meeting the demands of space missions, this invention provides a real-time tracking method and system for non-cooperative targets in space. This method ensures tracking performance even in obstructed environments and for targets of small size, achieving high-precision tracking of non-cooperative targets in various environments. This invention offers advantages such as high target tracking accuracy and strong anti-interference capabilities, enabling high-precision real-time tracking of obstructed and small-sized non-cooperative targets in space, thus improving the accuracy and robustness of non-cooperative target tracking.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] This invention discloses a real-time tracking method for non-cooperative targets in space, used for the detection and tracking of such targets. YOLOv5 is selected as the target detection algorithm model, and a lightweight MobileNetV3 model replaces CSPDarkent53 as the backbone network of the YOLOv5 target detection algorithm model, reducing the number of parameters and improving the detection speed. To further improve the detection speed, a lightweight PANet network is constructed, using depthwise separable convolutions instead of the traditional convolutional layers. Through multiple upsampling, downsampling, and feature fusion operations, the number of parameters in the YOLOv5 target detection algorithm is further reduced, improving the detection speed. The constructed lightweight YOLOv5 model is trained using a non-cooperative target database to obtain the optimal non-cooperative target detection model. The semi-major axis and eccentricity range of the spatial target are defined. Based on a single observation of the current non-cooperative target image information, the allowable domain method is used to determine the possible envelope of the spatial target location, denoted as Ω. x Where x represents the spatial target state vector. The region Ω where the spatial target's location may exist is defined. x Mapping this onto the camera plane yields the possible region corresponding to the bounding box of the target within the camera plane, denoted as Ω. X ,in For the state variables corresponding to the target bounding box, a Gaussian mixture distribution is used to represent the possible regions Ω corresponding to the target bounding box in the camera plane. X By approximating, the accuracy and robustness of selecting the target's bounding box region are improved. Through... A filtering dynamics model and an observation model are established using a seven-dimensional state vector. Based on these models, a Gaussian mixture filter algorithm is used iteratively to predict the target state in the next frame based on the detection results. A Hungarian algorithm is then used to correlate the detected and predicted targets. Through cascaded matching, the weight of repeatedly occurring targets is increased using the Hungarian algorithm, and unmatched detection boxes are matched with already detected boxes using the Intersection over Union (IOU) matching method. The parameters of the matched tracking boxes are updated using a filtering update formula, thus predicting the target tracking box for the next time step and achieving tracking of non-cooperative targets in space.

[0007] This invention discloses a real-time tracking method for non-cooperative targets in space, comprising the following steps:

[0008] Step 1: YOLOv5 is selected as the model for the object detection algorithm, and the lightweight MobileNetV3 model is used to replace CSPDarkent53 as the backbone network of the YOLOv5 object detection algorithm model. This reduces the number of parameters in the YOLOv5 object detection algorithm model and improves the detection speed. To further improve the detection speed of the YOLOv5 object detection algorithm, a lightweight PANet network is constructed. Depthwise separable convolutions are used to replace the traditional convolutional layers of the PANet network. Through multiple upsampling, downsampling, and feature fusion operations, the number of parameters in the YOLOv5 object detection algorithm is further reduced, thus improving the detection speed of the YOLOv5 object detection algorithm model.

[0009] Step 2: Label the spatial non-cooperative target dataset and divide the labeled images into training, validation, and test sets. Train the constructed lightweight YOLOv5 model using the spatial non-cooperative target database to obtain the optimal spatial non-cooperative target detection model.

[0010] Step 3: Define the semi-major axis and eccentricity range of the space target. Based on a single observation of the current non-cooperative space target image information, use the admissible region method to determine the possible envelope of the space target's location, denoted as Ω. x Where x represents the spatial target state vector. The region Ω where the spatial target's location may exist is defined. x Mapping this onto the camera plane yields the possible region corresponding to the bounding box of the target within the camera plane, denoted as Ω. X ,in Here are the state variables corresponding to the target bounding box: u is the x-coordinate of the center of the bounding box, v is the y-coordinate of the center of the bounding box, s is the area of ​​the bounding box, and r is the aspect ratio of the bounding box. It is the rate of change of the x-coordinate of the detection box center. It is the rate of change of the ordinate of the detection box center. It is the rate of change of the area of ​​the detection box, using a Gaussian mixture distribution to represent the possible regions Ω corresponding to the target box in the camera plane. X Approximation is used to improve the accuracy and robustness of selecting the target's bounding box region.

[0011] The mixture Gaussian distribution is represented as:

[0012]

[0013] Where Θ represents the mixture Gaussian distribution, p(X|Θ) is the probability of the mixture Gaussian distribution, and w k μ is the weight of the k-th sub-Gaussian distribution. k Let σ be the mean of the k-th sub-Gaussian distribution. kLet X be the covariance matrix of the k-th sub-Gaussian distribution, where K is a user-defined parameter used to characterize the number of sub-Gaussian distributions, and N = 7 represents the dimension of the state vector X.

[0014] Step Four: Through The seven-dimensional state vector is used to establish the filtering dynamics model and the observation model.

[0015] The following dynamic model is established:

[0016]

[0017] Where the subscripts i+1 and i represent t i+1 With t i The state vector at any given time.

[0018] The observation equations are established as follows:

[0019]

[0020] Z i+1 Indicates t i+1 The observation vector at time.

[0021] Step 5: Based on the dynamic model and observation model constructed in Step 4, an iterative Gaussian mixture filter algorithm is used to predict the state of the target in the next frame using the detection results. The Hungarian algorithm is then used to correlate the detected and predicted targets. Based on the dynamic model and observation model in Step 4, the Gaussian mixture filter algorithm is used iteratively to obtain estimated values ​​of the image after Gaussian mixture filtering. These estimates serve as initial values ​​for the next detection box prediction, including the predicted x-coordinate of the target box center point, the predicted y-coordinate of the target box, the predicted rate of change of the target box x-coordinate, the predicted rate of change of the target box y-coordinate, the predicted rate of change of the target box area, and the predicted aspect ratio of the target box.

[0022] Based on the dynamic model (2) and the observation model (3), the Gaussian mixture filter algorithm is used for iteration. The state of the target in the next frame is predicted by the detection results, and the data association between the detected target and the predicted target is realized by the Hungarian algorithm.

[0023] Make a prediction:

[0024]

[0025]

[0026] in and This represents the mean and covariance matrix of a one-step prediction.

[0027] Next, we will conduct observations and forecasts:

[0028]

[0029]

[0030]

[0031] in Let represent the observed prediction mean, observed covariance, and cross covariance, respectively, and R be the observed noise matrix.

[0032] Step six involves using a cascaded matching approach, based on the Hungarian algorithm, to increase the weight of targets that appear multiple times. The cosine and Mahalanobis distances of the detection and prediction results are calculated using the cosine and Mahalanobis distance formulas. Based on experimentally obtained thresholds, matching values ​​that do not conform to the Mahalanobis and cosine distances are removed. The parameters of the matched tracking boxes are updated using a filtering update formula. These updated parameters are then substituted into the dynamics and observation models to obtain the target tracking box for the next time step, thus achieving the tracking of non-cooperative targets in space.

[0033] Cosine distance calculation formula:

[0034]

[0035] Formula for calculating Mahalanobis distance:

[0036]

[0037] The Hungarian algorithm is used for matching to obtain matched and unmatched items. For unmatched detection boxes, the Intersection over Union (IOU) matching method is used to match them with the detected boxes. The parameters of the matched tracking boxes are updated using a filtering update formula.

[0038] Calculate the Kalman filter gain and update the parameters:

[0039]

[0040]

[0041]

[0042] Substituting the updated parameters into formulas (11) and (12) yields the target tracking box for the next moment, enabling the tracking of non-cooperative targets in space.

[0043] This invention also discloses a real-time tracking system for non-cooperative targets in space, used to implement the aforementioned real-time tracking method for non-cooperative targets in space. The real-time tracking system for non-cooperative targets in space includes a lightweight non-cooperative target detection module, a position extraction module, a detection box matching module, a Kalman filter module, and a storage module.

[0044] A lightweight spatial non-cooperative object detection module is trained on a large-scale dataset, and then fine-tuned based on a small number of spatial non-cooperative object sample images. The optimal lightweight object detection model after fine-tuning is used to obtain the location anchor boxes of non-cooperative objects, and the category and target region of the object to be tracked are determined based on the anchor box information. The category and target region information of the object to be tracked are input into the location extraction module of the lightweight spatial non-cooperative object detection module.

[0045] The location extraction module, based on the current spatial non-cooperative target category and target region information input by the lightweight spatial non-cooperative target detection module, uses the allowed domain method from step three to determine the possible envelope of the spatial target location.

[0046] The detection box matching module uses cascaded matching to increase the weight of frequently occurring targets. For trackers, detectors, and trackers in an uncertain state that still do not match after matching, the Intersection over Union (IOU) matching method is used to match the detection box with the trajectory.

[0047] The Kalman filter module iterates and updates the matched prediction boxes using the Kalman filter method to obtain a set of spatial non-cooperative target trajectories.

[0048] The storage module compares the current frame target estimation result of the non-cooperative target in space with the estimation results of the neighboring frames, and fits the multi-frame data into the motion trajectory of the non-cooperative target for storage, so as to save data for subsequent tasks.

[0049] Beneficial effects:

[0050] 1. The present invention discloses a real-time tracking method and system for non-cooperative targets in space. It uses lightweight non-cooperative target detection in space, uses MobileNetV3 as the backbone of the algorithm, and uses depthwise separable convolutions to replace the traditional convolutional layers of the PANet network. Then, through multiple upsampling, downsampling and feature fusion operations, the number of model parameters is further reduced and the detection speed of the model is improved.

[0051] 2. The present invention discloses a method and system for real-time tracking of non-cooperative targets in space. The initial distribution of the non-cooperative target boxes does not depend on human setting. The initial target boxes are obtained according to the allowable domain and are non-Gaussian distributed, which is closer to reality and the initial distribution is more accurate, thereby improving the robustness and convergence of target tracking.

[0052] 3. The present invention discloses a real-time tracking method and system for non-cooperative targets in space. Based on a single observation of the current image information of non-cooperative targets in space, the permissive domain method is used to determine the possible envelope of the target location, and the accuracy and robustness of the target location region selection are improved by using a mixture of Gaussian distributions.

[0053] 4. The present invention discloses a real-time tracking method and system for non-cooperative targets in space. It constructs a dynamic model and an observation model using a state vector. Then, based on the constructed model, it uses a Gaussian mixture filter algorithm for iteration. The state of the target in the next frame is predicted through the detection results. The Hungarian algorithm is used to realize the data association between the detected target and the predicted target. The weight of targets that appear multiple times is increased. The result after data association is used as the input of Kalman filter for further iteration, thereby improving the speed and accuracy of target tracking.

[0054] 5. The present invention discloses a real-time tracking method and system for non-cooperative targets in space. Through a lightweight target detection model and an improved target tracking algorithm, the speed and accuracy of target tracking are improved. This method is advantageous for deployment on spaceborne computers with limited computing power, thereby enhancing the real-time performance, accuracy, and robustness of tracking non-cooperative targets in space. Attached Figure Description

[0055] Figure 1 This is a flowchart of a real-time tracking method for non-cooperative targets in space according to the present invention;

[0056] Figure 2 This is a schematic diagram of the algorithm for the spatial non-cooperative target detection model in this invention;

[0057] Figure 3 This is a flowchart of the target tracking algorithm in this invention;

[0058] Figure 4 This is a schematic diagram of the real-time tracking system for non-cooperative targets in space in this invention. Detailed Implementation

[0059] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that the following embodiments are for illustrative purposes only and should not be considered as limiting the scope of the invention.

[0060] Reference Figure 1 This embodiment discloses a real-time tracking method and system for non-cooperative targets in space, and the specific implementation steps are as follows:

[0061] Step 1: As Figure 2As shown, YOLOv5 is selected as the object detection algorithm model, and the lightweight MobileNetV3 model is used to replace CSPDarkent53 as the backbone network of the YOLOv5 algorithm model, reducing the number of parameters and improving the detection speed of the model. MobileNetV3 first extracts image features through a traditional convolution, and then uses an inverse residual structure to stack them to achieve more refined feature extraction. At the same time, an SE attention mechanism is added to part of the inverse residual structure. The output vector is obtained through two fully connected layers. The number of nodes in the first fully connected layer is 1 / 4 of the number of channels in the feature matrix, and the number of nodes in the second fully connected layer is equal to the number of channels in the feature matrix. The output vector is derived by analyzing the weight relationship of each channel of the feature matrix, which improves the ability to extract features from the image. The activation function is redesigned and the swish function is replaced with h-swish.

[0062]

[0063] To further improve the detection speed of the algorithm, a lightweight PANet network was constructed. Depthwise separable convolutions were used to replace the traditional convolutional layers of the PANet network. Then, through multiple upsampling, downsampling, and feature fusion operations, the number of parameters of the algorithm was further reduced, thereby improving the detection speed of the model.

[0064] The GIOU loss function is introduced as a measure of anchor box loss. Compared to the traditional IOU, GIOU is scale-insensitive and considers not only overlapping areas but also other non-overlapping areas, thus better reflecting the degree of overlap. The formula for calculating GIOU is:

[0065]

[0066] Step 2: Using LabelImg as a labeling tool, import the images into LabelImg, mark the locations of spatially non-cooperative objects in the images with boxes, and indicate their type names. Complete the labeling of all images using this method, generating XML format label files, and output them to a specified folder to create a spatially non-cooperative object dataset. Then, use software to label the database, dividing it into training, validation, and test sets. Finally, train the constructed lightweight YOLOv5 model using the spatially non-cooperative object dataset to obtain the optimal spatially non-cooperative object detection model.

[0067] Step 3: Define the semi-major axis and eccentricity range of the space target. Based on a single observation of the current non-cooperative space target image information, use the admissible region method to determine the possible envelope of the space target's location, denoted as Ω. x Where x represents the spatial target state vector. The region Ω where the spatial target's location may exist is defined. xMapping this onto the camera plane yields the possible region corresponding to the bounding box of the target within the camera plane, denoted as Ω. X ,in For the state variables corresponding to the target bounding box, a Gaussian mixture distribution is used to represent the possible regions Ω corresponding to the target bounding box in the camera plane. X An approximation is performed to improve the accuracy and robustness of selecting the target's bounding box region. The Gaussian mixture distribution is represented as:

[0068]

[0069] Where Θ represents the mixture Gaussian distribution, p(X|Θ) is the probability of the mixture Gaussian distribution, and w k μ is the weight of the k-th sub-Gaussian distribution. k Let σ be the mean of the k-th sub-Gaussian distribution. k Let X be the covariance matrix of the k-th sub-Gaussian distribution, where K is a user-defined parameter used to characterize the number of sub-Gaussian distributions, and N = 7 represents the dimension of the state vector X.

[0070] Step Four: Through The seven-dimensional state vector is used to establish the filtering dynamics model and the observation model. The following dynamics model is established:

[0071]

[0072] Where the subscripts i+1 and i represent t i+1 With t i The state vector at any given time.

[0073] The observation equations are established as follows:

[0074]

[0075] Z i+1 Indicates t i+1 The observation vector at time.

[0076] Step 5, as Figure 3 As shown, based on the dynamic model and observation model constructed in step four, a Gaussian mixture filter algorithm is used iteratively to predict the state of the target in the next frame through the detection results, and the Hungarian algorithm is used to realize the data association between the detected target and the predicted target. Based on the dynamic model and observation model in step four, the Gaussian mixture filter algorithm is used iteratively to obtain the estimated values ​​of the image after Gaussian mixture filtering, which serve as the initial values ​​for the next detection box prediction. These include the predicted values ​​of the x-coordinate of the target box center point, the predicted values ​​of the target box y-coordinate, the predicted values ​​of the target box x-coordinate change rate, the predicted values ​​of the target box y-coordinate change rate, the predicted values ​​of the target box area change rate, and the predicted values ​​of the target box aspect ratio.

[0077] Based on the dynamic model (4) and the observation model (5), the Gaussian mixture filter algorithm is used for iteration. The state of the target in the next frame is predicted by the detection results, and the data association between the detected target and the predicted target is realized by the Hungarian algorithm.

[0078] First, let's make a prediction:

[0079]

[0080]

[0081] in and This represents the mean and covariance matrix of a one-step prediction.

[0082] Next, we will conduct observations and forecasts:

[0083]

[0084]

[0085]

[0086] in Let represent the observed prediction mean, observed covariance, and cross covariance, respectively, and R be the observed noise matrix.

[0087] Step six involves using cascaded matching, based on the Hungarian algorithm, to increase the weight of targets that appear multiple times. The cosine and Mahalanobis distances of the detection and prediction results are calculated using the cosine and Mahalanobis distance formulas. Then, based on thresholds obtained from experiments, matching values ​​that do not conform to the Mahalanobis and cosine distances are removed.

[0088] Cosine distance calculation formula:

[0089]

[0090] Formula for calculating Mahalanobis distance:

[0091]

[0092] The Hungarian algorithm is used for matching to obtain matching and non-matching items. For non-matching detection boxes, the Intersection over Union (IOU) matching method is applied. Finally, the parameters of the matching tracking boxes are updated using a filtering update formula.

[0093] Calculate the Kalman filter gain and update the parameters:

[0094]

[0095]

[0096]

[0097] Substituting the updated parameters into formulas (13) and (14) yields the target tracking box for the next moment, enabling the tracking of non-cooperative targets in space.

[0098] This invention also discloses a real-time tracking system for non-cooperative space targets, used to implement the aforementioned real-time tracking method for non-cooperative space targets. The real-time tracking system for non-cooperative space targets includes a lightweight non-cooperative space target detection module, a position extraction module, a detection box matching module, a Kalman filter module, and a storage module, such as... Figure 4 As shown.

[0099] A lightweight spatial non-cooperative object detection module is trained on a large-scale dataset, and then fine-tuned based on a small number of spatial non-cooperative object sample images. The optimal lightweight object detection model after fine-tuning is used to obtain the location anchor boxes of non-cooperative objects, and the category and target region of the object to be tracked are determined based on the anchor box information. The category and target region information of the object to be tracked are input into the location extraction module of the lightweight spatial non-cooperative object detection module.

[0100] The location extraction module, based on the current spatial non-cooperative target category and target region information input by the lightweight spatial non-cooperative target detection module, uses the allowed domain method from step three to determine the possible envelope of the spatial target location.

[0101] The detection box matching module uses cascaded matching to increase the weight of frequently occurring targets. For trackers, detectors, and trackers in an uncertain state that still do not match after matching, the Intersection over Union (IOU) matching method is used to match the detection box with the trajectory.

[0102] The Kalman filter module iterates and updates the matched prediction boxes using the Kalman filter method to obtain a set of spatial non-cooperative target trajectories.

[0103] The storage module compares the current frame target estimation result of the non-cooperative target in space with the estimation results of the neighboring frames, and fits the multi-frame data into the motion trajectory of the non-cooperative target for storage, so as to save data for subsequent tasks.

[0104] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is used to explain the present invention. It is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time tracking of non-cooperative targets in space, characterized in that: Includes the following steps, Step 1: Select YOLOv5 as the model for the object detection algorithm, and replace CSPDarkent53 with the lightweight MobileNetV3 model as the backbone network of the YOLOv5 object detection algorithm model to reduce the number of parameters and improve the detection speed. Construct a lightweight PANet network, and use depthwise separable convolutions to replace the traditional convolutional layers of the PANet network. Through multiple upsampling, downsampling, and feature fusion operations, further reduce the number of parameters of the YOLOv5 object detection algorithm and improve the detection speed. Step 2: Label the spatial non-cooperative target dataset and divide the labeled images into training, validation, and test sets; train the constructed lightweight YOLOv5 model using the spatial non-cooperative target database to obtain the optimal spatial non-cooperative target detection model; Step 3: Define the semi-major axis and eccentricity range of the space target. Based on a single observation of the current non-cooperative space target image information, use the admissible region method to determine the possible envelope of the space target's location, denoted as Ω. x , where x represents the spatial target state vector; the region Ω where the spatial target may exist. x Mapping this onto the camera plane yields the possible region corresponding to the bounding box of the target within the camera plane, denoted as Ω. X ,in Here are the state variables corresponding to the target bounding box: u is the x-coordinate of the center of the bounding box, v is the y-coordinate of the center of the bounding box, s is the area of ​​the bounding box, and r is the aspect ratio of the bounding box. It is the rate of change of the x-coordinate of the detection box center. It is the rate of change of the ordinate of the detection box center. It is the rate of change of the area of ​​the detection box, using a Gaussian mixture distribution to represent the possible regions Ω corresponding to the target box in the camera plane. X Approximation is used to improve the accuracy and robustness of selecting the target's bounding box region; Step 4: Establish the filtering dynamics model and observation model using the seven-dimensional state vector; Step 5: Based on the dynamic model and observation model constructed in Step 4, the Gaussian mixture filter algorithm is used for iteration. The state of the target in the next frame is predicted through the detection results, and the data association between the detected target and the predicted target is realized through the Hungarian algorithm. Based on the dynamic model and observation model in Step 4, the Gaussian mixture filter algorithm is used for iteration to obtain the estimated value of the image after Gaussian mixture filtering. This estimate is used as the initial value for the next detection box prediction, including the predicted value of the x-coordinate of the target box center point, the predicted value of the target box y-coordinate, the predicted value of the target box x-coordinate change rate, the predicted value of the target box y-coordinate change rate, the predicted value of the target box area change rate, and the predicted value of the target box aspect ratio. Step 6: Using a cascaded matching method, the weight of targets that appear multiple times is increased based on the Hungarian algorithm; The cosine distance and Mahalanobis distance of the detection and prediction results are calculated according to the cosine distance calculation formula and the Mahalanobis distance calculation formula. Based on the threshold obtained from the experiment, the matching values ​​that do not conform to the Mahalanobis distance and cosine distance are removed. The parameters of the matched tracking box are updated using a filtering update formula. The updated parameters are then substituted into the dynamic model and the observation model to obtain the target tracking box at the next moment, thus realizing the tracking of non-cooperative targets in space.

2. The method for real-time tracking of non-cooperative targets in space as described in claim 1, characterized in that: In step three, The mixture Gaussian distribution is represented as: Where Θ represents the mixture Gaussian distribution, p(X|Θ) is the probability of the mixture Gaussian distribution, and w k μ is the weight of the k-th sub-Gaussian distribution. k Let σ be the mean of the k-th sub-Gaussian distribution. k Let X be the covariance matrix of the k-th sub-Gaussian distribution, where K is a user-defined parameter used to characterize the number of sub-Gaussian distributions, and N = 7 represents the dimension of the state vector X.

3. The method for real-time tracking of non-cooperative targets in space as described in claim 2, characterized in that: In step four, The following dynamic model is established: Where the subscripts i+1 and i represent t i+1 With t i The state vector at any given time; The observation equations are established as follows: Z i+1 Indicates t i+1 The observation vector at time.

4. The method for real-time tracking of non-cooperative targets in space as described in claim 3, characterized in that: In step five, Based on the dynamic model (2) and the observation model (3), the Gaussian mixture filter algorithm is used for iteration, the state of the target in the next frame is predicted by the detection results, and the data association between the detected target and the predicted target is realized by the Hungarian algorithm. Make a prediction: in and This represents the mean and covariance matrix of a one-step prediction. Next, we will conduct observations and forecasts: in Let represent the observed prediction mean, observed covariance, and cross covariance, respectively, and R be the observed noise matrix.

5. The method for real-time tracking of non-cooperative targets in space as described in claim 4, characterized in that: In step six, Cosine distance calculation formula: Formula for calculating Mahalanobis distance: The Hungarian algorithm is used to match and obtain matching and non-matching items. For the unmatched detection boxes, the Intersection over Union (IOU) matching method is used to match them with the detected boxes. The parameters of the matched tracking boxes are updated using a filtering update formula. Calculate the Kalman filter gain and update the parameters: Substituting the updated parameters into formulas (11) and (12) yields the target tracking box for the next moment, enabling the tracking of non-cooperative targets in space.

6. A real-time tracking system for non-cooperative targets in space, used to implement a real-time tracking method for non-cooperative targets in space as described in claim 1, 2, 3, 4, or 5, characterized in that: It includes a lightweight spatial non-cooperative target detection module, a location extraction module, a detection box matching module, a Kalman filtering module, and a storage module; A lightweight spatial non-cooperative object detection module is trained on a large-scale dataset, and then the trained object detection model is fine-tuned based on a small number of spatial non-cooperative object sample images. The optimal lightweight target detection model with fine-tuning is used to obtain the location anchor boxes of non-cooperative targets. The category and target region of the object to be tracked are determined based on the anchor box information. The lightweight spatial non-cooperative target detection module extracts the category and target region information of the object to be tracked from the location extraction module. The location extraction module, based on the current spatial non-cooperative target category and target region information input by the lightweight spatial non-cooperative target detection module, determines the possible envelope of the spatial target location using the allowed domain method in step three. The detection box matching module uses cascaded matching to increase the weight of frequently occurring targets. For trackers, detectors, and trackers in an uncertain state that still do not match after matching, the Intersection over Union (IOU) matching method is used to match the detection box with the trajectory. The Kalman filter module iterates and updates the matched prediction boxes using the Kalman filter method to obtain a set of spatial non-cooperative target trajectories. The storage module compares the current frame target estimation result of the non-cooperative target in space with the estimation results of the neighboring frames, and fits the multi-frame data into the motion trajectory of the non-cooperative target for storage, so as to save data for subsequent tasks.

Images