GEO target detection method based on improved yolov8
Through the improved yolov8 deep detection network and sliding window method, the low signal-to-noise ratio and complex background problems of GEO target detection in low-cost ground telescopes are solved, and high-precision GEO target recognition and tracking are achieved, and the detection accuracy rate is more than 90%. .
Patent Information
- Application Number
- CN202410948926.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
It is difficult for the prior art to effectively identify and track high-value satellites and space debris in geosynchronous orbits, especially in low-cost ground telescope observations, which are difficult to detect targets due to factors such as low signal-to-noise ratio, cloud occlusion, atmospheric impact and light pollution.
Using the improved yolov8 deep detection network, by establishing a small object detection network yolov81_sod, adding a small detection head, training the image data set, and using the sliding window method to perform trajectory association and linear fitting of multi-frame images, eliminating false points, and filling them out Missing points improve the accuracy of target detection.
It has achieved high-precision recognition of GEO targets in low-cost telescope images, with a detection accuracy of more than 90%, and overcomes the problem of target recognition in low signal-to-noise ratio and complex background environments.
Smart Images

Figure CN118918312A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of space situational awareness in the field of aerospace, and relates to a technology for observing GEO targets through a low-cost ground telescope, and in particular to a GEO target detection method based on improved YOLO V8. Background Art
[0002] Space target detection and tracking is the basis of space situational awareness. There are a large number of high-value satellites deployed in the geosynchronous orbit, including communication, navigation, meteorological and other satellites. In order to protect the safety of these space assets, it is particularly important to monitor the geosynchronous orbit.
[0003] Detecting space targets through star maps is an important application in astronomy and space science, which involves analyzing and processing star map images obtained from telescopes or other observation equipment to identify and track space targets including satellites and space debris. Although traditional large-aperture telescopes have low noise and high resolution, they are also relatively expensive. Therefore, it is of positive significance to use low-cost, low-precision ground telescope cameras to monitor large airspace for a long time.
[0004] Affected by the rotation of the earth and the exposure time of the camera, star targets appear as stripes in the image, while GEO targets appear as dots in the image because their movement speed is almost synchronized with that of ground equipment. Although GEO targets can be distinguished from star backgrounds by their shapes, due to factors such as long observation distance, low image resolution, cloud and star occlusion, atmospheric influence, light pollution, and sensor noise, GEO targets are usually dim in energy and have low signal-to-noise ratios. They are easily lost in the star and noise background and blocked by clouds, which makes it difficult to accurately identify targets.
[0005] Therefore, the existing monitoring methods need to be improved. Summary of the invention
[0006] Purpose of the invention: In order to overcome the above shortcomings, the purpose of the present invention is to provide a GEO target detection method based on improved yolov8, which adopts a deep detection network to improve the efficiency and accuracy of target detection, and solves the problem that the image taken by the low-cost telescope camera on the ground has a low signal-to-noise ratio, and the target detection is difficult due to factors such as clouds, star occlusion, atmospheric influence, light pollution and sensor noise.
[0007] Technical solution: In order to achieve the above purpose, the present invention provides a GEO target detection method based on improved yolov8, comprising the following steps:
[0008] S1): Establish a small target detection network yolov81_sod based on yolov81, that is, the original yolov81 has three detection heads, whose sizes are 32×32, 16×16, and 8×8 targets respectively. On this basis, a small detection head is added, whose size is 4×4, which is specially used to detect small targets;
[0009] S2): Prepare a data set to train the network, that is, the background noise distribution of the training set image is close to the actual image, the GEO target position is marked in the training sample, and the number of GEO targets contained in each frame image in the training sample is 0, 1 or more, and the data label is a 4×4, or 6×6, or 8×8 rectangular area. The target is located in the center of the area. The small target detection network based on yolov81 established in S1) is used for training;
[0010] S3): Use the telescope camera to take a sequence of GEO target images, collect images at equal intervals during the shooting process, and keep the camera still during the sampling period;
[0011] S4): Input the captured image into the network to obtain the coordinates of the candidate points, that is, the coordinates of 0 or more candidate points in the image can be obtained from the network prediction results;
[0012] S5): Perform trajectory association on candidate points of multiple frames of images, remove false points, add missing points, and obtain the GEO target position.
[0013] The GEO target detection method based on improved yolov8 described in the present invention, the specific process of S2): preparing a data set to train the network is as follows:
[0014] S201): Add output head to yolov81 and establish small target detection network model yolov81_sod;
[0015] S202): Convert the training data labels in the target data set into YOLO format, where the data label format is a 6×6 rectangle, and the coordinate system corresponding to the original data label is the center of the rectangle;
[0016] S203): Use the network model yolov81_sod to train the data set;
[0017] S204): Verify and evaluate the accuracy of the test set data.
[0018] The GEO target detection method based on improved yolov8 described in the present invention, the specific process of S5): performing trajectory association and straight line fitting on candidate points of multiple frames of continuous images, eliminating false points, and supplementing missing points is as follows:
[0019] S501) Use a sliding window for target tracking. Assume the window size is m, where m >= 5. At time t, there are n actual targets in the tracking window. Each target in each frame of the image is represented as O i,j , where i represents the image frame number, i = 1,..., m; j represents the target number, j = 1,..., n;
[0020] O i,j represents the j-th target in frame i, and the total number of targets in frame i is n; while the candidate targets obtained through network detection are O' i,k , and the number of detected targets in frame i is l. If l > n, after this step, the extra false targets will be eliminated; if l < n, after this step, the missed targets will be supplemented;
[0021] S502): For the point targets of geostationary satellites, the points of each target on each frame of the image form a straight line. If the images are taken at equal intervals, the distances between the same target in adjacent two frames are almost equal;
[0022] S503): All points not on the candidate straight line are regarded as false targets and are eliminated;
[0023] S504): If the number of candidate points in the candidate straight line is less than the maximum number of frames, candidate points are supplemented in the frames lacking candidate points, and the coordinate positions of the supplemented points are calculated based on the known distances between points in adjacent two frames;
[0024] S505): The sliding window method can detect newly emerging targets in the middle of the observation period.
[0025] For the GEO target detection method based on the improved yolov8 described in the present invention, the S52): For the point targets of geostationary satellites, the points of each target on each frame of the image form a straight line. When performing linear fitting, the straight line is represented by the linear equation ax + by + c:
[0026] First, find all the straight lines formed by two points as alternative straight lines, and then traverse each candidate point in the remaining image frames one by one to calculate whether it is collinear with a certain alternative straight line;
[0027] If they are collinear, check whether the difference in intervals between adjacent points is less than the specified threshold. If the detection passes, add it to the alternative straight line point set. After screening all the points, delete the alternative straight lines with less than 3 data points to ensure that the final obtained candidate straight line sequence contains at least 3 or more candidate points.
[0028] From the above technical solutions, it can be seen that the present invention has the following beneficial effects:
[0029] The GEO target detection method based on improved Yolov8 described in the present invention solves the problem of low accuracy of detection of small targets by the yolo network by adding the yolov81 output head, establishes a detection network yolov81_sod suitable for small targets to train a training data set composed of continuous multi-frame images, so that the network can identify multiple point targets in the image with high accuracy. The point targets identified by the network are used as candidate targets, and by establishing a sliding window, the candidate points contained in the image frame in the window are track-associated and straight-line-fitted, false targets are removed, and targets that are not identified by the network due to factors such as cloud occlusion and complex background are supplemented to obtain the final GEO target. The method is implemented on the spotGEO data set, and the detection accuracy can reach more than 90%. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 It is a flow chart of the GEO target detection method based on improved Yolov8 involved in the present invention;
[0031] Figure 2 The association relationship of small and weak targets in multiple frames of sliding window images;
[0032] Figure 3 Training results of small object detection network on spotGEO data;
[0033] Figure 4 The recognition accuracy of the trained weak target detection network on the validation dataset;
[0034] Figure 5 Comparison of network detection results and trajectory association results. DETAILED DESCRIPTION
[0035] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments.
[0036] Example 1
[0037] As shown in the figure, a GEO target detection method based on improved yolov8 includes the following steps:
[0038] S1): Establish a small target detection network yolov81_sod based on yolov81, that is, the original yolov81 has three detection heads, whose sizes are 32×32, 16×16, and 8×8 targets respectively. On this basis, a small detection head is added, whose size is 4×4, which is specially used to detect small targets;
[0039] S2): Prepare a data set to train the network, that is, the background noise distribution of the training set image is close to the actual image, the GEO target position is marked in the training sample, and the number of GEO targets contained in each frame image in the training sample is 0, 1 or more, and the data label is a 4×4, or 6×6, or 8×8 rectangular area. The target is located in the center of the area. Use yolov81_sod established in S1) for training;
[0040] S3): Use the telescope camera to take a sequence of GEO target images, collect images at equal intervals during the shooting process, and keep the camera still during the sampling period;
[0041] S4): Input the captured image into the network to obtain the coordinates of the candidate points, that is, the coordinates of 0 or more candidate points in the image can be obtained from the network prediction results;
[0042] S5): Perform trajectory association and straight line fitting on candidate points of multiple frames of continuous images, remove false points, and add missing points.
[0043] In the GEO target detection method based on improved yolov8 described in this embodiment, the specific process of S2): preparing a data set to train the network is as follows:
[0044] S201): Add output head to yolov81 and establish small target detection network model yolov81_sod;
[0045] S202): Convert the labels of the training data in the data set (a total of 1280 groups, each with 5 frames) into the YOLO format. The data label format is a 6×6 rectangle. The coordinate system corresponding to the original data label is the center of the rectangle. The data set is randomly divided, 80% of which is divided into a training set and 20% into a validation set.
[0046] S203): Use the network model yolov81_sod to train the data set;
[0047] S204): Verify and evaluate the accuracy of the test set data.
[0048] In this embodiment, the specific process of S5): performing trajectory association and straight line fitting on candidate points of multiple frames of continuous images, eliminating false points, and supplementing missing points is as follows:
[0049] S501) Use sliding window to track the target, such as Figure 2 As shown, assuming that the window size is m, m>=5; there are actually n targets in the tracking window at time t, and the target in each frame is represented by O i,j , where i represents the image frame number, i=1,...,m; j represents the target number, j=1,...,n;
[0050] Oi,j Denote the j-th target in frame i, and the total number of targets in frame i is n; the candidate targets obtained through network detection are O'. i,k , the number of targets detected in frame i is l. If l > n, after this step, the extra false targets will be removed; if l < n, after this step, the missed targets will be supplemented;
[0051] S502): For each point target of the geostationary satellite, the points of each target on each frame image form a straight line. If the images are taken at equal intervals, the distances between the same target in adjacent two frames are almost equal; when performing linear fitting, the straight line is represented by the linear equation ax + by + c. First, find all the straight lines formed by two points as candidate straight lines, and then traverse the candidate points in the remaining image frames one by one to calculate whether they are collinear with a certain candidate straight line;
[0052] If they are collinear, check whether the interval difference between adjacent points is less than the specified threshold. If the detection passes, add the points to the candidate straight line point set. After screening all the points, delete the candidate straight lines with less than 3 data points to ensure that the final obtained candidate straight line sequence contains at least 3 candidate points;
[0053] S503): All the points not on the candidate straight lines are regarded as false targets and are removed;
[0054] S504): If the number of candidate points in the candidate straight line is less than the maximum number of frames, supplement candidate points in the frames lacking candidate points, and calculate the coordinate positions of the supplemented points according to the known distances between points in adjacent two frames;
[0055] S505): The sliding window method can be used to detect the newly emerged targets in the middle of the observation time period.
[0056] Embodiment 2
[0057] The following is a specific description with actual data:
[0058] A GEO target detection method based on improved Yolov8 described in this embodiment establishes a weak target detection network by improving yolov8 to identify candidate point targets in telescope images, and determines the final GEO targets through the association of point target trajectories in multiple frames of images. This method is applicable to detecting GEO targets from ground telescope images.
[0059] As shown in the figure, a GEO target detection method based on improved yolov8 includes the following steps:
[0060] S1): Establish a small target detection network yolov81_sod based on yolov81, that is, the original yolov81 has three detection heads, whose sizes are 32×32, 16×16, and 8×8 targets respectively. On this basis, a small detection head is added, whose size is 4×4, which is specially used to detect small targets;
[0061] S2): Prepare a data set to train the network, that is, the background noise distribution of the training set image is close to the actual image, the GEO target position is marked in the training sample, and the number of GEO targets contained in each frame image in the training sample is 0, 1 or more, and the data label is a 4×4, or 6×6, or 8×8 rectangular area. The target is located in the center of the area. The small target detection network based on yolov81 established in S1) is used for training;
[0062] S3): Use the telescope camera to take a sequence of GEO target images, collect images at equal intervals during the shooting process, and keep the camera still during the sampling period;
[0063] S4): Input the captured image into the network to obtain the coordinates of the candidate points, that is, the coordinates of 0 or more candidate points in the image can be obtained from the network prediction results;
[0064] S5): Perform trajectory association and straight line fitting on candidate points of multiple frames of images, remove false points, and add missing points.
[0065] The method of the present invention is specifically described with reference to the embodiments:
[0066] In this embodiment:
[0067] S201): Add output head to the small target detection network model based on yolov81;
[0068] S202): The training data (a total of 1280 groups, each with 5 frames) in the spotGEO data set released by ESA are converted into yolo format. The data label format is a 6×6 rectangle. The coordinate system corresponding to the original data label is the center of the rectangle. The data set is randomly divided, 80% of which is divided into a training set and 20% is a validation set. S203): The data set is trained through a network model. The network model training parameters are set as follows:
[0069] Number of images per batch batch = 8
[0070] Initial learning rate lr0 = 0.01
[0071] Final learning rate lrf = 0.01
[0072] Momentum of SGD optimizer = 0.937
[0073] Optimizer weight decay weight_decay = 0.0005
[0074] The number of rounds in the warm-up phase warmup_epochs = 3.0
[0075] Initial momentum of the warm-up phase warmup_momentum = 0.8
[0076] The initial bias learning rate in the warm-up phase is warmup_bias_lr = 0.1;
[0077] The training results are shown in Figure 3 , the recognition accuracy on the validation dataset is shown in Figure 4 ;
[0078] S204): Verify using the test set provided by spotGEO. The test set data has 5120 groups, each with 5 frames of images.
[0079] S205): Input the test set data into the network, obtain the prediction result label file, and put it into the specified folder;
[0080] S206): Evaluate the prediction results to obtain F1, and the evaluation criteria are defined as follows: TP represents True positive, N TP Indicates the number of targets that are accurately identified within a certain error range, N FP Indicates false positive, N FP Indicates the number of other objects misidentified as targets, FN stands for false negative, N FN Represents the number of missed targets, from which the calculation formulas for accuracy Precision and recall Recall can be defined:
[0081]
[0082]
[0083]
[0084] The result obtained in S204) is evaluated and F1=0.843 is obtained;
[0085] The GEO target detection method based on improved yolov8 described in this embodiment, the specific process of S5): performing trajectory association and straight line fitting on candidate points of multiple frames of images, eliminating false points, and supplementing missing points is as follows:
[0086] S501) uses a sliding window to track the target. The principle is as follows Figure 2As shown, assume the window size is m, where m >= 5; at time t, there are n targets actually in the tracking window, and each target in each frame of the image is represented as O i,j , where i represents the image frame sequence number, i = 1,..., m; j represents the target sequence number, j = 1,..., n;
[0087] O i,j represents the j-th target in frame i, and the total number of targets in frame i is n; while the candidate targets obtained through network detection are O' i,k , the number of detected targets in frame i is l. If l > n, after this step, the extra false targets will be removed; if l < n, after this step, the missed targets will be supplemented;
[0088] S502): For the point targets of geostationary satellites, the points of each target on each frame of the image form a straight line. If the images are taken at equal intervals, the distances between the same target in adjacent two frames are almost equal; when performing linear fitting, the straight line is represented by the linear equation ax + by + c. First, find all the straight lines formed by two points as the alternative straight lines, and then traverse the candidate points in the remaining image frames one by one to calculate whether they are collinear with a certain alternative straight line;
[0089] If they are collinear, check whether the interval difference between adjacent points is less than the specified threshold. If the detection passes, add them to the alternative straight line point set. After screening all the points, delete the alternative straight lines with less than 3 data points to ensure that the final obtained candidate straight line sequence contains at least 3 or more candidate points;
[0090] S503): All the points not on the candidate straight lines are regarded as false targets and are removed;
[0091] S504): If the number of candidate points in the candidate straight line is less than the maximum number of frames, supplement candidate points in the frames lacking candidate points, and calculate the coordinate positions of the supplemented points according to the known distances between points in adjacent two frames;
[0092] S505): The sliding window method can be used to detect the newly emerging targets in the middle of the observation period.
[0093] The effects before and after linear fitting are shown in Figure 5 in Figure 5In Figure 1, three typical cases are shown. The left picture shows the candidate points predicted directly by the network, and the right picture shows the point targets after screening and supplementation. In order to intuitively show the effect, the results of the detection of the five frames are superimposed on the middle frame picture. The five colors of green, blue, red, purple and yellow represent the point targets in the 1st to 5th frames. It can be seen that the first row (corresponding to the picture sequence with the test set number 100) shows the effect of simply supplementing the missing point targets; the second row (corresponding to the picture sequence with the test set number 178) supplements the first target at the 2nd and 3rd frame positions (the points in the 5th frame are beyond the image boundary and cannot be displayed), the third target is supplemented at the 4th frame position, the first target is supplemented at the 2nd and 3rd frame positions, and a false target in the 3rd frame is deleted. In the third row (corresponding to the picture sequence with the test set number 1002), the first target in the figure detects the complete sequence, and the points in the 1st and 2nd frames are beyond the image boundary and cannot be displayed. After trajectory association, the third target was added in the first frame, the fourth target was added in the fourth frame, and the two false targets in the fifth frame were deleted.
[0094] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.
Claims
1. A GEO target detection method based on improved yolov8, characterized in that: The following steps are involved: S1): Establish a small target detection network yolov81_sod based on yolov81, that is, the original yolov81 has three detection heads, whose sizes are 32×32, 16×16, and 8×8 targets respectively. On this basis, a small detection head is added, whose size is 4×4, which is specially used to detect small targets; S2): Prepare a data set to train the network, that is, the background noise distribution of the training set image is close to the actual image, the GEO target position is marked in the training sample, and the number of GEO targets contained in each frame image in the training sample is 0, 1 or more, and the data label is a 4×4, or 6×6, or 8×8 rectangular area. The target is located in the center of the area. Use yolov81_sod established in S1) for training; S3): Use the telescope camera to take a sequence of GEO target images, collect images at equal intervals during the shooting process, and keep the camera still during the sampling period; S4): Input the captured image into the network to obtain the coordinates of the candidate points, that is, the coordinates of 0 or more candidate points in the image can be obtained from the network prediction results; S5): Perform trajectory association and straight line fitting on candidate points of multiple consecutive image frames, remove false points, add missing points, and obtain the GEO target position.
2. The GEO target detection method based on improved yolov8 according to claim 1, characterized in that: The specific process of S2): preparing the data set to train the network is as follows: S201): Add output head to yolov81 and establish small target detection network model yolov81_sod; S202): converting the target data label into a yolo format, the data label format is a 6×6 rectangle, and the coordinate system corresponding to the original data label is the center of the rectangle; S203): Use the network model yolov81_sod to train the data set; S204): Verify and evaluate the accuracy of the test set data.
3. The GEO target detection method based on improved yolov8 according to claim 1, characterized in that: The specific process of S5): performing trajectory association and straight line fitting on candidate points of multiple frames of continuous images, eliminating false points, and supplementing missing points is as follows: S501) Use sliding window to track the target, assuming that the window size is m, m>=5; there are actually n targets in the tracking window at time t, and the target in each frame is represented by O i,j , where i represents the image frame number, i=1,...,m; j represents the target number, j=1,...,n; O i,j represents the j-th target in frame i, and the total number of targets in frame i is n; while the candidate target obtained through network detection is O' i,k , the number of targets detected in frame i is l. If l > n, after this step, the extra false targets will be removed; if l < n, after this step, the undetected targets will be supplemented; S502): For geosynchronous satellite point targets, each target point on each frame image forms a straight line. If the images are taken at equal intervals, the distance between two adjacent frames of the same target is almost equal; S503): All points not on the candidate straight line are regarded as false targets and are eliminated; S504): If the number of candidate points in the candidate line is less than the maximum number of frames, then supplement the candidate points in the frames where the candidate points are missing, and the coordinate positions of the supplementary points are calculated based on the distances between the known points in two adjacent frames; S505): The sliding window method can be used to detect new targets that appear in the middle of the observation time period.
4. The GEO target detection method based on improved yolov8 according to claim 3, characterized in that: S502): For geosynchronous satellite point targets, each point of each target on each frame image forms a straight line. When performing straight line fitting, the straight line is represented by a straight line equation ax+by+c: First, find all the lines formed by two points as candidate lines, and then traverse the candidate points in the remaining image frames one by one to calculate whether they are collinear with a candidate line; If they are collinear, check whether the interval difference between adjacent points is less than the specified threshold. If the detection is passed, add them to the candidate line point set. After all points are screened, delete the candidate lines with less than 3 data points to ensure that the final candidate line sequence contains at least 3 candidate points.
Citation Information
Patent Citations
Dark and dim point target detection method in infrared image
CN107330857A
GEO space target extraction method based on deep learning
CN114782527A
Moving point target time-space domain trajectory tracking method based on line segment association
CN116958189A
Crack monitoring method and terminal
CN117523385A
Cited By
Space target vision measurement method combined with acceleration drift estimation
CN119879947A
A space target visual measurement method combined with acceleration drift estimation
CN119879947B
