An anti-interference rotating target tracking method and an embedded device

Abstract

The present invention relates to a target tracking method and an embedded device, and particularly to an anti-interference rotating target tracking method and an embedded device, which solve the problems that the existing methods cannot accurately track rotating targets and have complex calculations. The method includes the following steps: Step 1: Obtain the first frame of image block; Step 2: Affine rotation; Step 3: Obtain the rotation angle α of the target in the second frame; Step 4: Obtain the final position of the target in the second frame; Step 5: Obtain the second frame as the current frame; Step 6: Obtain the current frame image block; Step 7: Affine rotation; Step 8: Obtain the rotation angle α of the target in the next frame; Step 9: Judge; If not occluded, execute Step 10, if occluded, execute Step 11; Step 10: Update the parameters with the response map position as the final position of the target in the next frame and execute Step 12; Step 11: Fuse the Kalman filter to obtain the final position of the target in the next frame; Step 12: Judge; If not ended, update the current frame and return to Step 6.

Description

Technical Field

[0001] The present invention relates to a target tracking method and an embedded device, and particularly to an anti-interference rotating target tracking method and an embedded device. Background Art

[0002] Tracking ground targets by remote sensing satellites has important application values in disaster detection, traffic monitoring, marine protection, etc.

[0003] When tracking ground targets by remote sensing satellites, there are many existing tracking methods available, but most of them cannot achieve accurate tracking of rotating targets. Although the Siamese R-CNN tracking method (a neural network tracking method for visual tracking through re-detection) can achieve relatively accurate tracking of rotating targets, during the tracking process, it requires a large amount of data training and the calculation process is very complex, making it impossible to be installed and run on an embedded device. However, due to the bird's-eye view of remote sensing observation, it is very common for the target to be tracked to be occluded and rotated. Therefore, there is an urgent need to develop a rotating target tracking method with a simple calculation process, which can be installed and run on an embedded device and can accurately track targets with occlusion and rotation. Summary of the Invention

[0004] The object of the present invention is to solve the technical problem that when most existing tracking methods track targets with occlusion and rotation, they cannot achieve accurate tracking, and some can achieve relatively accurate tracking, but the calculation process is too complex to be installed and run on an embedded device, and to provide an anti-interference rotating target tracking method and an embedded device.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is:

[0006] An anti-interference rotating target tracking method, characterized in that it includes the following steps:

[0007] Step 1: Obtain the first frame of image, frame the position of the target to be tracked on this image, and obtain the image block of the target to be tracked on the first frame of image;

[0008] Step 2: Perform affine rotation on the image block obtained in Step 1 for respectively set n pairs of positive and negative symmetric angles to obtain 2n rotated image blocks, where n is a natural number greater than 1;

[0009] Step 3: Based on the image patches obtained in Step 1 and the 2n rotated image patches obtained in Step 2, use the kernel correlation filtering model to obtain the maximum target response map of the second frame image; then, according to the affine rotation angle of one of the rotated image patches corresponding to the maximum target response map of the second frame image among the affine rotation angles when obtaining the 2n rotated image patches in Step 2, obtain the predicted rotation angle α of the target to be tracked in the second frame image; the rotation angle α of the target to be tracked refers to the rotation angle of the target to be tracked relative to the target to be tracked in its previous frame image;

[0010] Step 4: Take the position corresponding to the maximum target response map of the second frame image obtained in Step 3 as the final position of the target to be tracked in the predicted second frame image, and based on this final position, update the parameters of the kernel correlation filtering model in Step 3;

[0011] Step 5: Obtain the second frame image as the current frame image;

[0012] Step 6: Based on the final position of the target to be tracked in the current frame image predicted from the previous frame image, obtain the image patch of the target to be tracked on the current frame image;

[0013] Step 7: Respectively perform affine rotation on the image patches obtained in Step 6 based on the rotation angle α of the target to be tracked in the current frame image predicted from the previous frame image and the nearest in-pool angle β corresponding to this angle α; the nearest in-pool angle β is the angle value among the n pairs of positive and negative symmetric angles set in Step 2 that is closest to the angle α and has an absolute value greater than the absolute value of the angle α. When the angle α is equal to the angle with the largest absolute value among the n pairs of positive and negative symmetric angles set in Step 2, the nearest in-pool angle β takes the angle value that is closest to the angle α among them except for the angle α;

[0014] Step 8: Based on the image patches obtained in Step 6 and the two rotated image patches obtained in Step 7, use the kernel correlation filtering model with the parameters updated most recently to obtain the maximum target response map of the next frame image; then, according to the affine rotation angle of one of the rotated image patches corresponding to the maximum target response map of the next frame image among the affine rotation angles when obtaining the two rotated image patches in Step 7, obtain the predicted rotation angle α of the target to be tracked in the next frame image;

[0015] Step 9: According to the preset occlusion judgment criterion for the target to be tracked, judge whether the target to be tracked is occluded; if not, execute Step 10; if so, execute Step 11;

[0016] Step 10: Take the position corresponding to the maximum target response map of the next frame image obtained in Step 8 as the final position of the target to be tracked in the predicted next frame image, and based on this final position, update the parameters of the kernel correlation filtering model last used in Step 8; then execute Step 12;

[0017] Step 11: According to the maximum target response map of the next frame image obtained in Step 8 and the motion displacements of the target to be tracked in the recent M frames, fuse the Kalman filter to obtain the final position of the target to be tracked in the predicted next frame image, where M is an integer greater than or equal to 3; if the total number of frames of the images before the current frame image is less than M frames, then take the motion displacements of the target to be tracked from the first frame image to the current frame image, fuse the Kalman filter, and perform the acquisition;

[0018] Step 12: Determine whether an end tracking instruction is received; if not, obtain the next frame image as the new current frame image, update the current frame image, and return to Step 6; if so, complete the tracking.

[0019] Further, in Step 1, the position of the target to be tracked is framed by manually framing or introducing a lightweight deep learning neural network for detecting the target to be tracked.

[0020] Further, in Step 1 and Step 6, the area of the image patch of the target to be tracked obtained is 2.5 to 3.5 times the area of the target to be tracked, and its specific value is determined by the size of the target to be tracked and is an empirical value;

[0021] In Step 2, the absolute values of the set n pairs of positive and negative symmetric angles are all greater than zero and less than or equal to 3 degrees.

[0022] Further, in order to comprehensively cover the rotation angle of the target to be tracked between frames and make the calculation process simple and the image affine rotation effect obvious, in Step 2, n is less than or equal to 4; the set n pairs of positive and negative symmetric angles are selected from ±0.5 degrees, ±1 degree, ±1.5 degrees, and ±2 degrees.

[0023] Further, in Step 3, the specific process of obtaining the rotation angle α of the target to be tracked in the predicted second frame image is: among the affine rotation angles of the 2n rotated image patches obtained in Step 2, take the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the second frame image directly as the rotation angle α of the target to be tracked in the obtained predicted second frame image;

[0024] In step 8, the specific process of obtaining the rotation angle α of the target to be tracked in the predicted next frame image is as follows: Among the affine rotation angles when obtaining the two rotated image patches in step 7, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the next frame image is directly used as the rotation angle α of the target to be tracked in the predicted next frame image obtained.

[0025] Alternatively, in order to make the predicted rotation angle α of the target to be tracked more accurate and thus achieve more accurate tracking of the rotating target, in step 3, the specific process of obtaining the rotation angle α of the target to be tracked in the predicted second frame image is as follows:

[0026] Among the affine rotation angles when obtaining 2n rotated image patches in step 2, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the second frame image is used as the theoretical value of the rotation angle α of the target to be tracked in the predicted second frame image obtained. Then, the scale-invariant feature transform algorithm is used to measure the rotation angle α of the target to be tracked, and the measurement result is used as the measured value of the rotation angle α of the target to be tracked in the predicted second frame image obtained. Then, the average of the theoretical value and the measured value is taken as the rotation angle α of the target to be tracked in the predicted second frame image obtained.

[0027] In step 8, the specific process of obtaining the rotation angle α of the target to be tracked in the predicted next frame image is as follows:

[0028] Among the affine rotation angles when obtaining the two rotated image patches in step 7, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the next frame image is used as the theoretical value of the rotation angle α of the target to be tracked in the predicted next frame image obtained. Then, the scale-invariant feature transform algorithm is used to measure the rotation angle α of the target to be tracked, and the measurement result is used as the measured value of the rotation angle α of the target to be tracked in the predicted next frame image obtained. Then, the average of the theoretical value and the measured value is taken as the rotation angle α of the target to be tracked in the predicted next frame image obtained.

[0029] Furthermore, in step 9, the specific method for determining whether the target to be tracked is occluded according to the preset occlusion judgment criterion for the target to be tracked is as follows:

[0030] It is determined according to the magnitude relationship between the peak value of the maximum target response map of the next frame image obtained in step 8 and the set occlusion detection threshold: When the peak value of the maximum target response map of the next frame image obtained in step 8 is greater than the set occlusion detection threshold, it is determined as no, and the target to be tracked is not occluded; otherwise, it is determined as yes, and the target to be tracked is occluded.

[0031] The set occlusion detection threshold is taken as 0.3 - 0.5, and its specific value is determined by the size and motion characteristics of the target to be tracked, which is an empirical value.

[0032] Alternatively, in step 9, according to the preset occlusion judgment criterion for the target to be tracked, the specific method for judging whether the target to be tracked is occluded is as follows:

[0033] A lightweight deep learning neural network is introduced to detect the target to be tracked, and a judgment is made based on the detection result: if the target to be tracked can be detected and the confidence level is greater than 0.5, it is judged as no, and the target to be tracked is not occluded; otherwise, it is judged as yes, and the target to be tracked is occluded.

[0034] Furthermore, in order to make the final position of the target to be tracked in the predicted next frame of image more accurate when the target to be tracked is occluded, and thus achieve more accurate tracking, and at the same time make the calculation process relatively simple, in step 11, M is taken as 3 - 5, and its specific value is determined by the motion characteristics of the target to be tracked, which is an empirical value.

[0035] Meanwhile, the present invention also provides an embedded device, including a memory, a processor, and a computer program stored on the memory; the special feature is that: the processor executes the computer program to implement the steps of the above anti-interference rotating target tracking method.

[0036] The beneficial effects of the present invention are:

[0037] (1)The anti-interference rotating target tracking method of the present invention performs affine rotation of a plurality of pairs of positive and negative symmetric angles on the image block of the target to be tracked in the first frame image to obtain a plurality of rotated image blocks. According to the image block before rotation and the plurality of rotated image blocks, a kernel correlation filtering model is used to obtain the maximum target response map of the second frame image, and then the rotation angle α of the target to be tracked in the predicted second frame image and the final position of the target to be tracked in the predicted second frame image are obtained, and the parameters of the kernel correlation filtering model are updated; then, for the image block of the target to be tracked in the subsequent frame images, affine rotation is performed based on the rotation angle α of the target to be tracked in the current frame image predicted from the previous frame image and the nearest pooling angle β corresponding to the angle α, to obtain two rotated image blocks. According to the image block before rotation and the two rotated image blocks, the kernel correlation filtering model with the parameters updated most recently is used to obtain the maximum target response map of the next frame image, and then after obtaining the rotation angle α of the target to be tracked in the predicted next frame image, first determine whether the target to be tracked is occluded, and then according to the judgment result, use different methods to obtain the final position of the target to be tracked in the predicted next frame image, and when it is judged that no occlusion occurs, update the parameters of the kernel correlation filtering model, and when it is judged that occlusion occurs, do not update the parameters of the kernel correlation filtering model, and update only when it is judged that no occlusion occurs;

[0038] As can be seen from the above description of the acquisition process: The present invention can not only obtain the final position of the target to be tracked in the predicted next frame image, but also obtain the rotation angle α of the target to be tracked in the predicted next frame image, so it can achieve accurate tracking of the rotating target and improve the robustness of the kernel correlation filtering model; in addition, for subsequent frame images other than the first frame, during the process of obtaining the final position of the target to be tracked in the predicted next frame image, a judgment on whether the target to be tracked is occluded is introduced, and different methods are used to obtain the final position of the target to be tracked in the predicted next frame image for the case where the target to be tracked is not occluded and the case where the target to be tracked is occluded. Therefore, the final position of the target to be tracked in the predicted next frame image obtained by using the anti-interference rotating target tracking method of the present invention is more accurate, and the accuracy of tracking the occluded and rotating target is improved; moreover, when using the anti-interference rotating target tracking method of the present invention to obtain the final position of the target to be tracked in the predicted next frame image, its calculation process does not require a large amount of data training, the calculation process is simple, the requirements for hardware are low, and it can be installed and run on embedded devices;

[0039] In summary, the present invention solves the technical problems that most of the existing tracking methods cannot achieve accurate tracking when tracking occluded and rotating targets, and some can achieve relatively accurate tracking, but the calculation process is too complex and cannot be installed and run on embedded devices.

[0040] (2) In the anti-interference rotating target tracking method of the present invention, the n pairs of positive and negative symmetric angles set in step 2 are preferably selected from ±0.5 degrees, ±1 degree, ±1.5 degrees, and ±2 degrees. With such a setting, it can not only comprehensively cover the angles of rotation of the target to be tracked, but also make the calculation process simple, the image affine rotation effect obvious, and improve the accuracy of tracking.

[0041] (3) In the anti-interference rotating target tracking method of the present invention, when obtaining the rotation angle α of the target to be tracked in the predicted second frame image in step 3 and the rotation angle α of the target to be tracked in the predicted next frame image in step 8, it preferably further includes the step of measuring the rotation angle α of the target to be tracked by using the scale-invariant feature transform algorithm. In this way, the finally obtained rotation angle α of the target to be tracked is more accurate, and thus the accuracy of tracking can be improved.

[0042] (4) In the anti-interference rotating target tracking method of the present invention, the judgment of whether the target to be tracked is occluded is introduced, and preferably two judgment criteria are provided, which improves the anti-interference performance in the tracking process, and thus improves the accuracy of tracking the occluded and rotating target.

[0043] (5) When using the anti-interference rotating target tracking method of the present invention to track a rotating target, through experimental verification, its tracking accuracy can reach more than 96.7%. Description of the Drawings

[0044] Figure 1 is a flowchart of an embodiment of the anti-interference rotating target tracking method of the present invention;

[0045] Figure 2 is a schematic diagram of the tracking effect on the corresponding frame image when using the embodiment of the anti-interference rotating target tracking method of the present invention to track an airplane; wherein:

[0046] (a) is the first frame image;

[0047] (b) is the 50th frame image;

[0048] (c) is the 100th frame image;

[0049] (d) is the 150th frame image;

[0050] Figure 3 is a schematic diagram of the tracking effect on the corresponding frame image when using the embodiment of the anti-interference rotating target tracking method of the present invention to track a ship; wherein:

[0051] (e) is the first frame image;

[0052] (f) is the 249th frame image;

[0053] (g) is the 307th frame image;

[0054] (h) is the 326th frame image;

[0055] Figure 4 It is a comparison schematic diagram of the accuracy curve graphs for tracking a rotating target using different tracking methods. Specific embodiments

[0056] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] See Figure 1 , an anti-interference rotating target tracking method of the present invention includes the following steps:

[0058] Step 1: Obtain the first frame image, frame the position of the target to be tracked on this image, and obtain the image block of the target to be tracked on the first frame image;

[0059] When framing the position of the target to be tracked above, it can be framed manually, or other methods such as introducing lightweight deep learning neural networks such as YOLO_X (a variant of the YOLO series (You Only Look Once) target detection model, proposed by the Alibaba Cloud team and Megvii in 2021), YOLOv3 (the third version of the YOLO series (You Only Look Once) target detection model), and SSD (a single-stage target detection model) for detecting the target to be tracked can be used; in this embodiment, it is framed manually; in order to improve the accuracy and stability of tracking, when obtaining the image block of the target to be tracked on the first frame image above, the area of the obtained image block of the target to be tracked is preferably usually 2.5 to 3.5 times the area of the target to be tracked, and its specific value is determined by the size of the target to be tracked, which is an empirical value; if the target to be tracked is an aircraft or ship target, it is usually taken as 3 times the size; if the target to be tracked is a vehicle target, it is usually taken as 2.5 times the size; in this embodiment, the target to be tracked is an aircraft, and 3 times the area of the target to be tracked is taken;

[0060] Step 2: Perform affine rotations of the set of n pairs of positive and negative symmetric angles on the image block obtained in Step 1 respectively to obtain 2n rotated image blocks, where n is a natural number greater than 1;

[0061] The absolute values of the n pairs of positive and negative symmetric angles set above are preferably generally greater than zero and less than or equal to 3 degrees; in order to comprehensively cover the angles of inter-frame rotation of the target to be tracked and make the calculation process simple and the image affine rotation effect obvious, preferably n is less than or equal to 4, and the n pairs of positive and negative symmetric angles set above are selected from ±0.5 degrees, ±1 degrees, ±1.5 degrees, and ±2 degrees; for example, when n is 2, the n pairs of positive and negative symmetric angles can be ±0.5 degrees and ±1 degree; in this embodiment, in order to achieve higher tracking accuracy, preferably n is 4, that is, in this embodiment, the image blocks obtained in step 1 are respectively subjected to affine rotation of ±0.5 degrees, ±1 degree, ±1.5 degrees, and ±2 degrees to obtain 8 rotated image blocks;

[0062] Step 3: According to the image blocks obtained in step 1 and the 2n rotated image blocks obtained in step 2, use the kernel correlation filtering model to obtain the maximum target response map of the second frame image; then according to the affine rotation angle of one of the rotated image blocks corresponding to the maximum target response map of the second frame image among the affine rotation angles when obtaining the 2n rotated image blocks in step 2, obtain the predicted rotation angle α of the target to be tracked in the second frame image; the rotation angle α of the target to be tracked refers to the rotation angle of the target to be tracked relative to the target to be tracked in its previous frame image;

[0063] The specific process of obtaining the rotation angle α of the target to be tracked in the predicted second-frame image can be as follows: In step 2, among the affine rotation angles when obtaining 2n rotated image patches, the affine rotation angle of one rotated image patch corresponding to the maximum target response map of the second-frame image is directly used as the rotation angle α of the target to be tracked in the predicted second-frame image. For example, if one rotated image patch corresponding to the maximum target response map of the second-frame image is the image patch obtained by performing an affine rotation of 0.5 degrees on the image patch obtained in step 1 in step 2, then the rotation angle α of the target to be tracked in the predicted second-frame image obtained is 0.5 degrees. In addition to the above direct acquisition, it is also possible to first use the affine rotation angle of one rotated image patch corresponding to the maximum target response map of the second-frame image among the affine rotation angles when obtaining 2n rotated image patches in step 2 as the theoretical value of the rotation angle α of the target to be tracked in the predicted second-frame image, and then use the scale-invariant feature transform algorithm to measure the rotation angle α of the target to be tracked, use the measurement result as the measured value of the rotation angle α of the target to be tracked in the predicted second-frame image, and then take the average of the theoretical value and the measured value as the rotation angle α of the target to be tracked in the predicted second-frame image. Which method to specifically use to obtain the rotation angle α of the target to be tracked in the predicted second-frame image is determined according to the requirements for tracking accuracy in actual applications. In this embodiment, the affine rotation angle of one rotated image patch corresponding to the maximum target response map of the second-frame image among the affine rotation angles when obtaining 2n rotated image patches in step 2 is directly used as the rotation angle α of the target to be tracked in the predicted second-frame image, and the rotation angle α of the target to be tracked in the predicted second-frame image obtained in this embodiment is 0.5 degrees;

[0064] Step 4: Use the position corresponding to the maximum target response map of the second-frame image obtained in step 3 as the final position of the target to be tracked in the predicted second-frame image, and based on this final position, update the parameters of the kernel correlation filtering model in step 3;

[0065] Step 5: Obtain the second-frame image as the current-frame image;

[0066] Step 6: Based on the final position of the target to be tracked in the predicted current-frame image from the previous-frame image, obtain the image patch of the target to be tracked on the current-frame image;

[0067] When obtaining the image patch of the target to be tracked on the current frame image, similar to step 1, the area of the obtained image patch of the target to be tracked is preferably usually 2.5 to 3.5 times the area of the target to be tracked, and its specific value is determined by the size of the target to be tracked, which is an empirical value; if the target to be tracked is an aircraft or a ship target, usually take 3 times the size; if the target to be tracked is a vehicle target, usually take 2.5 times the size; in this embodiment, the target to be tracked is an aircraft, and take 3 times the area of the target to be tracked.

[0068] Step 7: For the image patches obtained in step 6, perform affine rotation on the rotation angle α of the target to be tracked in the current frame image predicted based on the previous frame image, and the nearest pool angle β corresponding to the angle α, respectively, to obtain two rotated image patches; the above-mentioned nearest pool angle β is the angle value among the n pairs of positive and negative symmetric angles set in step 2 that is closest to the angle α and has an absolute value greater than the absolute value of the angle α except the angle α. When the angle α is equal to the angle with the largest absolute value among the n pairs of positive and negative symmetric angles set in step 2, then the nearest pool angle β takes the angle value that is closest to the angle α except the angle α; in this embodiment, the nearest pool angle β is the angle value among ±0.5 degrees, ±1 degree, ±1.5 degrees, and ±2 degrees that is closest to the angle α and has an absolute value greater than the absolute value of the angle α except the angle α. When the angle α is equal to the angle with the largest absolute value among the n pairs of positive and negative symmetric angles set in step 2, then the nearest pool angle β takes the angle value that is closest to the angle α except the angle α; for example: when the angle α is equal to +1 degree, then the nearest pool angle β takes +1.5 degrees; when the angle α is equal to +2 degrees, then the nearest pool angle β takes +1.5 degrees; when the angle α is equal to -2 degrees, then the nearest pool angle β takes -1.5 degrees; when the current frame image is the second frame image, in this embodiment, the rotation angle α of the target to be tracked predicted in the second frame image obtained in step 3 above is 0.5 degrees, then the nearest pool angle β takes 1 degree. In this step 7, that is, perform affine rotation of 0.5 degrees and 1 degree respectively on the image patch of the target to be tracked on the second frame image obtained in step 6 to obtain two rotated image patches.

[0069] Step 8: According to the image patches obtained in step 6 and the two rotated image patches obtained in step 7, use the kernel correlation filtering model with the most recently updated parameters to obtain the maximum target response map of the next frame image; then according to the affine rotation angle of one of the rotated image patches corresponding to the maximum target response map of the next frame image among the affine rotation angles when obtaining the two rotated image patches in step 7, obtain the rotation angle α of the target to be tracked predicted in the next frame image.

[0070] The specific process of obtaining the rotation angle α of the target to be tracked in the predicted next-frame image is the same as that in Step 3, and can be: In Step 7, among the affine rotation angles when obtaining the two rotated image patches, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the next-frame image is directly used as the rotation angle α of the target to be tracked in the predicted next-frame image; or first, among the affine rotation angles when obtaining the two rotated image patches in Step 7, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the next-frame image is used as the theoretical value of the rotation angle α of the target to be tracked in the predicted next-frame image, and then the Scale-Invariant Feature Transform (SIFT) algorithm is used to measure the rotation angle α of the target to be tracked, and the measurement result is used as the measured value of the rotation angle α of the target to be tracked in the predicted next-frame image. Then, the average of the theoretical value and the measured value is used as the rotation angle α of the target to be tracked in the predicted next-frame image; which method is specifically selected to obtain the rotation angle α of the target to be tracked in the predicted next-frame image is determined according to the requirements for tracking accuracy in practical applications; in this embodiment, the affine rotation angle of the rotated image patch corresponding to the maximum target response map of the next-frame image among the affine rotation angles when obtaining the two rotated image patches in Step 7 is directly used as the rotation angle α of the target to be tracked in the predicted next-frame image; when the current frame image is the second frame image, the rotation angle α of the target to be tracked in the predicted next-frame image (i.e., in the third frame image) obtained in this embodiment is 1 degree;

[0071] Step 9: Determine whether the target to be tracked is occluded according to the preset occlusion judgment criterion for the target to be tracked; if not, execute Step 10; if so, execute Step 11;

[0072] In this embodiment, according to the preset occlusion judgment criterion for the target to be tracked, the specific method for judging whether the target to be tracked is occluded is as follows: judge according to the size relationship between the peak value of the maximum target response map of the next frame of image obtained in step 8 and the set occlusion detection threshold: when the peak value of the maximum target response map of the next frame of image obtained in step 8 is greater than the set occlusion detection threshold, it is judged as no, and the target to be tracked is not occluded; otherwise, it is judged as yes, and the target to be tracked is occluded. To improve the accuracy of judging the occlusion of the target to be tracked, the above-mentioned set occlusion detection threshold is preferably taken as 0.3-0.5, and its specific value is determined by the size and motion characteristics of the target to be tracked, which is an empirical value. Usually, when the target to be tracked is an aircraft target, the set occlusion detection threshold is taken as 0.3; when the target to be tracked is a ship target, the set occlusion detection threshold is taken as 0.35; when the target to be tracked is a vehicle target or other targets other than aircraft and ships, the set occlusion detection threshold is taken as 0.4. In this embodiment, the target to be tracked is an aircraft, and the set occlusion detection threshold is taken as 0.3. The above-mentioned specific method for judging whether the target to be tracked is occluded according to the preset occlusion judgment criterion for the target to be tracked, in addition to using the method of this embodiment, can also introduce lightweight deep learning neural networks such as YOLO_X, YOLOv3, and SSD to detect the target to be tracked, and judge according to the detection result: if the target to be tracked can be detected and the confidence level is greater than 0.5, it is judged as no, and the target to be tracked is not occluded; otherwise, it is judged as yes, and the target to be tracked is occluded.

[0073] Step 10: Take the position corresponding to the maximum target response map of the next frame of image obtained in step 8 as the final position of the target to be tracked in the predicted next frame of image, and based on this final position, update the parameters of the kernel correlation filtering model used most recently in step 8; then execute step 12.

[0074] Step 11: According to the maximum target response map of the next frame of image obtained in step 8 and the motion displacements of the target to be tracked in the most recent M frames, fuse the Kalman filter to obtain the final position of the target to be tracked in the predicted next frame of image, where the above-mentioned M is an integer greater than or equal to 3; if the total number of frames of the images before the current frame of image is less than M frames, then take the motion displacements of the target to be tracked from the first frame of image to the current frame of image, fuse the Kalman filter, and perform the acquisition.

[0075] In order to make the final position of the target to be tracked in the predicted next frame of image more accurate when the target to be tracked is occluded, so as to achieve more accurate tracking, and at the same time make the calculation process relatively simple, the above-mentioned M is preferably usually taken as 3-5, and its specific value is determined by the motion characteristics of the target to be tracked, which is an empirical value. Usually, when the target to be tracked is an aircraft or ship target, M is taken as 3, and when the target to be tracked is a vehicle, M is taken as 5. In this embodiment, M is taken as 3.

[0076] Step 12: Determine whether an end tracking instruction is received; if not, obtain the next frame of image as the new current frame image, update the current frame image, and then return to Step 6; if so, complete the tracking.

[0077] In addition, the present invention also provides an embedded device, including a memory, a processor, and a computer program stored on the memory; the processor executes the computer program to implement the steps of the above anti-interference rotating target tracking method.

[0078] Figure 2 It is a schematic diagram of the tracking effect on the corresponding frame image when the anti-interference rotating target tracking method of the present invention is used to track an aircraft; wherein: (a) is the first frame image; (b) is the 50th frame image; (c) is the 100th frame image; (d) is the 150th frame image. Figure 3 It is a schematic diagram of the tracking effect on the corresponding frame image when the anti-interference rotating target tracking method of the present invention is used to track a ship; wherein: (e) is the first frame image; (f) is the 249th frame image; (g) is the 307th frame image; (h) is the 326th frame image. Figure 2 and Figure 3 As can be seen from and , the anti-interference rotating target tracking method of the present invention can achieve accurate tracking of aircraft and ships, can effectively track rotating targets, estimate the target position when occlusion occurs, and maintain tracking of the target when occlusion ends.

[0079] Figure 4 It is a schematic comparison diagram of the accuracy curves of tracking rotating targets using different tracking methods. Figure 4 In , the accuracies of tracking using BOOSTING (an ensemble statistical learning tracking algorithm), MIL (multiple instance learning tracking algorithm), MEDIANFLOW (median optical flow tracking algorithm), Siamese R-CNN (a neural network tracking method for visual tracking through re-detection), ECO (efficient continuous convolution operator tracking algorithm), KCF (kernel correlation filtering algorithm), and RACFME (the anti-interference rotating target tracking method of the present invention) are compared, and their tracking accuracies are 47.6%, 42.1%, 8.2%, 90.5%, 77.8%, 73.4%, and 96.7% respectively. As can be seen from this, among these methods, the anti-interference rotating target tracking method of the present invention has the highest tracking accuracy for rotating targets.

[0080] In summary, the anti-interference rotating target tracking method of the present invention improves the anti-interference performance during the tracking process and can accurately track occluded and rotating targets with high precision.

Claims

1. A method for tracking a rotating target with interference resistance, characterized in that: The following steps are involved: Step 1: Obtain a first frame image, frame the position of the target to be tracked on the image, and obtain an image block of the target to be tracked on the first frame image; Step 2: Performing affine rotations of n pairs of positive and negative symmetric angles on the image blocks obtained in step 1 to obtain 2n rotated image blocks, where n is a natural number greater than 1; Step 3: According to the image block obtained in step 1 and the 2n rotated image blocks obtained in step 2, a maximum target response map of the second frame image is obtained by using a kernel correlation filter model; then, according to the image block corresponding to the maximum target response map of the second frame image in the 2n rotated image blocks in step 2, the corresponding affine rotation angle in step 2 is used to obtain the predicted rotation angle α of the target to be tracked in the second frame image; the rotation angle α of the target to be tracked refers to the rotation angle of the target to be tracked relative to the target to be tracked in the previous frame image; Step 4: taking the position corresponding to the maximum target response map of the second frame image obtained in step 3 as the predicted final position of the target to be tracked in the second frame image, and updating the parameters of the kernel correlation filter model in step 3 based on the final position; Step 5: Obtain the second frame image as the current frame image; Step 6: based on the final position of the target to be tracked in the current frame image predicted by the previous frame image, an image block of the target to be tracked on the current frame image is obtained; Step 7: For the image block obtained in step 6, perform affine rotation of the angle α of the target to be tracked in the current frame image predicted based on the previous frame image, and the angle β in the nearest pool corresponding to the angle α, respectively, to obtain two rotated image blocks; the angle β in the nearest pool is the angle value closest to the angle α and having an absolute value greater than the absolute value of the angle α among the n pairs of positive and negative symmetric angles set in step 2, except the angle α. When the angle α is equal to the angle with the largest absolute value among the n pairs of positive and negative symmetric angles set in step 2, the angle β in the nearest pool is the angle value closest to the angle α except the angle α. Step 8: According to the image block obtained in step 6 and the two rotated image blocks obtained in step 7, the maximum target response map of the next frame image is obtained by using the kernel correlation filter model whose parameters have been updated most recently; and then, according to the image block corresponding to the maximum target response map of the next frame image in the two rotated image blocks in step 7, the corresponding affine rotation angle in step 7 is used to obtain the predicted rotation angle α of the target to be tracked in the next frame image; Step 9: judging whether the target to be tracked is blocked according to the preset target to be tracked blockage judgment standard; If not, proceed to step 10; If yes, go to step 11; Step 10: taking the position corresponding to the maximum target response map of the next frame image obtained in step 8 as the predicted final position of the target to be tracked in the next frame image, and based on the final position, updating the parameters of the kernel correlation filter model most recently used in step 8; then executing step 12; Step 11: Based on the maximum target response map of the next frame image obtained in step 8 and the motion displacement of the target to be tracked in the most recent M frames, the Kalman filter is integrated to obtain the predicted final position of the target to be tracked in the next frame image, where M is an integer greater than or equal to 3; if the total number of frames before the current frame image is less than M frames, the motion displacement of the target to be tracked from the first frame image to the current frame image is obtained by integrating the Kalman filter; Step 12: Determine whether an end tracking instruction is received; if not, obtain the next frame image as a new current frame image, update the current frame image, and return to step 6; If yes, the tracking is completed.

2. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 1, the position of the target to be tracked is framed by manually framing the position of the target to be tracked or by introducing a lightweight deep learning neural network to detect the target to be tracked.

3. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 1 and step 6, the area of ​​the image block of the target to be tracked is 2.5 to 3.5 times the area of ​​the target to be tracked, and its specific value is determined by the size of the target to be tracked and is an empirical value; In step 2, the absolute values ​​of the set n pairs of positive and negative symmetry angles are all greater than zero and less than or equal to 3 degrees.

4. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 2, n is less than or equal to 4; the set n pairs of positive and negative symmetry angles are selected from ±0.5 degrees, ±1 degree, ±1.5 degrees and ±2 degrees.

5. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 3, the specific process of obtaining the predicted rotation angle α of the target to be tracked in the second frame image is as follows: the image block corresponding to the maximum target response map of the second frame image among the 2n rotated image blocks in step 2, the corresponding affine rotation angle in step 2, is directly used as the obtained predicted rotation angle α of the target to be tracked in the second frame image; In step 8, the specific process of obtaining the predicted rotation angle α of the target to be tracked in the next frame image is: the image block corresponding to the maximum target response map of the next frame image in the two rotated image blocks in step 7, the corresponding affine rotation angle in step 7, is directly used as the obtained predicted rotation angle α of the target to be tracked in the next frame image.

6. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 3, the specific process of obtaining the predicted rotation angle α of the target to be tracked in the second frame image is: The image block corresponding to the maximum target response map of the second frame image among the 2n rotated image blocks in step 2, the corresponding affine rotation angle in step 2, is used as the theoretical value of the rotation angle α of the target to be tracked in the predicted second frame image, and then the scale-invariant feature transformation algorithm is used to measure the rotation angle α of the target to be tracked, and the measurement result is used as the measurement value of the rotation angle α of the target to be tracked in the predicted second frame image, and then the theoretical value and the measured value are averaged to obtain the rotation angle α of the target to be tracked in the predicted second frame image; In step 8, the specific process of obtaining the predicted rotation angle α of the target to be tracked in the next frame image is: The image block corresponding to the maximum target response map of the next frame image in the two rotated image blocks in step 7, and the corresponding affine rotation angle in step 7 are used as the theoretical value of the rotation angle α of the target to be tracked in the predicted next frame image, and then the scale-invariant feature transformation algorithm is used to measure the rotation angle α of the target to be tracked, and the measurement result is used as the measured value of the rotation angle α of the target to be tracked in the predicted next frame image, and then the theoretical value and the measured value are averaged as the rotation angle α of the target to be tracked in the predicted next frame image.

7. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 9, according to the preset occlusion judgment standard of the target to be tracked, the specific method for judging whether the target to be tracked is occluded is: The judgment is made according to the magnitude relationship between the peak value of the maximum target response map of the next frame image obtained in step 8 and the set occlusion detection threshold: when the peak value of the maximum target response map of the next frame image obtained in step 8 is greater than the set occlusion detection threshold, it is judged as no, and the target to be tracked is not occluded; otherwise, it is judged as yes, and the target to be tracked is occluded; The set occlusion detection threshold is 0.3-0.5, and its specific value is determined by the size and motion characteristics of the target to be tracked, and is an empirical value.

8. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 9, according to the preset occlusion judgment standard of the target to be tracked, the specific method for judging whether the target to be tracked is occluded is: A lightweight deep learning neural network is introduced to detect the target to be tracked, and a judgment is made based on the detection results: if the target to be tracked can be detected and the confidence level is greater than 0.5, it is judged as no, and the target to be tracked is not occluded; otherwise, it is judged as yes, and the target to be tracked is occluded.

9. The anti-interference rotating target tracking method according to claim 1, characterized in that: In step 11, M is 3-5, and its specific value is determined by the motion characteristics of the target to be tracked and is an empirical value.

10. An embedded device, comprising a memory, a processor, and a computer program stored in the memory; characterized in that: The processor executes the computer program to implement the steps of the interference-resistant rotating target tracking method according to any one of claims 1 to 9.

Images