High-robustness target loss detection and rapid tracking method

Abstract

The invention discloses a high-robustness target loss detection and rapid tracking method, and belongs to the technical field of calculation, reckoning or counting. By combining the YOLO algorithm and the MOSSE algorithm, the method can actively search the target and realize re-detection and rapid recovery tracking of the target after the space target is lost for a long time. And when the target moves out of the visual field, the MOSSE algorithm cannot continue tracking, and at the moment, the YOLO algorithm is automatically started to carry out target detection so as to confirm that the target appears again. Once a target is detected, the features of the target are immediately transmitted to the MOSSE algorithm so as to quickly switch back to the tracking mode, thereby ensuring the continuity and stability of tracking. The key point of the method is that a switching mechanism can quickly respond between target loss and reoccurrence, and time delay in the detection and tracking process is remarkably reduced.

Description

Technical Field

[0001] The invention relates to target tracking technology, and in particular discloses a highly robust target loss detection and rapid tracking method, belonging to the technical field of calculation, estimation or counting. Background Art

[0002] With the rapid development of science and technology and the continuous advancement of space technology, outer space has gradually become a new field of scientific and military competition. As an important branch of space technology research, target tracking is mainly responsible for detecting and tracking space targets in Earth orbit to determine their key parameters such as type, attitude and trajectory. This technology not only helps to gain a deeper understanding of the dynamic changes in outer space, but also plays an important role in future space attack and defense.

[0003] At present, traditional target tracking methods for natural celestial bodies such as stars and planets are already very mature and can achieve multiple effects such as precise positioning and continuous tracking. However, for near-Earth space targets, such as artificial satellites, their motion patterns are significantly different from those of stellar celestial bodies. These targets not only have faster angular velocities, but may also change speed and orbit during movement, or even move out of the field of view, leading to complex situations such as target loss. At this time, if the system relies on a traditional single tracking method, it may not be able to resume tracking in time after the target is lost, resulting in continuity interruption and inability to accurately track near-Earth space targets such as artificial satellites. Summary of the invention

[0004] In view of the deficiencies of the above-mentioned background technology, the present invention provides a highly robust target loss detection and rapid tracking method to solve the technical problem that a space target cannot be tracked after it is lost for a long time and then reappears, thereby achieving the purpose of the invention of re-detecting a space target and quickly resuming tracking after the space target is lost for a long time.

[0005] The present invention adopts the following technical solutions to achieve the above-mentioned invention object:

[0006] A highly robust target loss detection and fast tracking method comprises the following steps:

[0007] Step 1: Initialize the space target to a lost state, enter the detection mode, call the YOLO algorithm to extract the feature information of the space target, and obtain the target detection result;

[0008] Step 2: Enter the tracking mode, initialize the MOSSE algorithm according to the target detection result, pre-train the filter generated by the initialized MOSSE algorithm, and obtain the target tracking result;

[0009] Step three: learn the difference between target detection results and target tracking results through the reward mechanism, and switch between detection mode and tracking mode according to the learning results of the reward mechanism.

[0010] As a further optimization scheme for a highly robust target loss detection and fast tracking method, step one calls the YOLO algorithm to extract spatial target feature information, which specifically includes the following steps:

[0011] Step 1-1, annotate and normalize the space target image, use the satellite model to describe the space target, construct the space target image including diversified data of different angles and different lighting conditions, train the space target image of diversified data, and obtain the space target image data set training model;

[0012] Step 1-2, extract features from the input space target image based on the model trained by the space target image dataset.

[0013] As a further optimization scheme for a robust target loss detection and fast tracking method, step 2 pre-trains the filter generated by the MOSSE algorithm. When the filter is initialized for the first time, the number of pre-training iterations is set to the maximum value. When the filter is initialized for the subsequent time, the number of pre-training iterations is set to 1. The process of pre-training single-step iteration includes the following steps:

[0014] Step 2-1, convert the current frame target image into the frequency domain, and apply the filter generated by the initialized MOSSE algorithm to perform convolution operation, calculate the response value that characterizes the degree of matching between each position and the target feature, obtain the current frame response map, and take the position corresponding to the maximum response value as the estimated position of the spatial target;

[0015] Step 2-2: when the maximum response value is lower than the set threshold, proceed to step 2-3; otherwise, return to step 2-1;

[0016] Step 2-3, based on the current response value of the filter and the current target detection result obtained by the YOLO algorithm, the filter is adaptively updated, the target position is updated, and the target tracking result is output.

[0017] As a further optimization scheme for a highly robust target loss detection and fast tracking method, step three learns the difference between the target detection result and the target tracking result through a reward mechanism based on a reinforcement learning algorithm: when the difference between the target detection result and the target tracking result is less than a set threshold, a positive reward for switching to the tracking mode is output; when the difference between the target detection result and the target tracking result is greater than or equal to the set threshold, a negative reward for switching to the detection mode is output.

[0018] As a further optimization scheme for a highly robust target loss detection and fast tracking method, the specific method of step 1-1 to construct a space target image including diverse data of different angles and different lighting conditions is: performing enhancement operations on the space target image including but not limited to image rotation, image scaling, and brightness adjustment.

[0019] As a further optimization scheme for a highly robust target loss detection and fast tracking method, steps 1-2 extract features from the input space target image based on the space target image dataset training model. Specifically, feature layers of different scales are obtained through a feature pyramid structure, predictions including bounding boxes, category probabilities, and confidence scores are generated on feature layers of different scales, bounding boxes and predictions with confidence scores lower than a set threshold are discarded, overlapping bounding boxes are eliminated through a non-maximum suppression algorithm, and target detection results are output.

[0020] As a further optimization scheme for a robust target loss detection and fast tracking method, image rotation rotates a given space target image point (x, y) by an angle θ to obtain the rotated space target image point (x′, y′). A positive value of θ indicates counterclockwise rotation, and a negative value of θ indicates clockwise rotation.

[0021] As a further optimization scheme for a robust target loss detection and fast tracking method, image scaling adjusts a given spatial target image point (x, y) by a scaling factor s to obtain a scaled spatial target image point (x′, y′). s>1 means the image is enlarged, and 0<s<1 means the image is reduced.

[0022] As a further optimization scheme for a highly robust target loss detection and fast tracking method, brightness adjustment is used to adjust the contrast and brightness of the original space target image I to obtain the adjusted space target image I′, where I′=I·α+β, α is the contrast factor, and the value range of α is [0.8, 2.0], and β is the brightness offset, and the value range of β is [-100, 100].

[0023] As a further optimization scheme for a robust target loss detection and fast tracking method, steps 2-3 adaptively update the filter based on the current response value of the filter and the current target detection result obtained by the YOLO algorithm. The expression is: Ht +1 =(1-α)Ht+αFt, where H t+1 is the response value at time t+1, H t is the response value at time t, F tis the target feature information detected by the YOLO algorithm at time t, α is the learning rate, when the difference between the maximum response value in the current frame response map and the maximum response value in the previous frame response map is greater than the set threshold, the value of α is 0.7-0.9, when the difference between the maximum response value in the current frame response map and the maximum response value in the previous frame response map is less than or equal to the set threshold, the value of α is 0.3-0.5.

[0024] The present invention adopts the above technical solution and has the following beneficial effects:

[0025] (1) According to the method of the present invention, by combining the reinforcement learning mechanism, rapid detection and recovery tracking after the loss of a target that moves rapidly and changes orbits frequently in a space environment is achieved. In particular, in a highly dynamically changing orbital environment, reinforcement learning can analyze the output difference between YOLO detection and MOSSE tracking in real time, adjust the detection and tracking strategy, and achieve the purpose of rapid switching to cope with satellite attitude changes and orbit interference, ensuring efficient continuous tracking in this environment and reducing tracking interruptions caused by target loss.

[0026] (2) According to the method of the present invention, by expanding and diversifying the data set of the YOLO algorithm, the characteristics of space targets are optimized in particular. By introducing different lighting conditions, different speed conditions and simulation of satellite attitude changes, it is ensured that the algorithm can accurately identify and track high-speed moving targets in orbit in practical applications, thereby achieving more stable tracking performance.

[0027] (3) According to the method of the present invention, by reducing the number of iterations of the MOSSE algorithm and introducing an adaptive update strategy, the response speed and accuracy of the algorithm in space target tracking are significantly improved. This improvement enables the algorithm to quickly adapt to the rapid orbit change and speed change of the target during the satellite orbit motion, improves the tracking sensitivity to dynamic targets, and ensures stable and effective execution of tracking tasks in a space environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Shown is an overall flow chart of the target loss detection and rapid tracking method proposed in the present invention.

[0029] Figure 2 Shown is a flow chart of the YOLO algorithm in the present invention for detecting space targets.

[0030] Figure 3 Shown is a flow chart of the MOSSE algorithm for tracking space targets in the present invention. DETAILED DESCRIPTION

[0031] The technical solution of the invention is described in detail below with reference to the accompanying drawings.

[0032] In view of the technical problem that the space target cannot be tracked back in time after being lost for a long time, the present invention provides a more stable and reliable target tracking method, which can actively search for the target and resume tracking after the space target is lost for a long time. The method combines the YOLO algorithm with the MOSSE algorithm to achieve rapid switching between detection and tracking, ensuring the continuity and stability of target tracking. This combination not only improves the algorithm's adaptability to dynamic changes in the target, but also significantly reduces the probability of tracking failure caused by target loss, thereby improving the overall robustness.

[0033] like Figure 1 As shown, a highly robust target loss detection and fast tracking method proposed by the present invention includes steps one to three.

[0034] Step 1: Call the YOLO algorithm for target detection

[0035] In the specific implementation process, the target state is initially set to "lost". In this state, the detection mode is entered and the YOLO algorithm is called to perform target loss detection. When the target is successfully detected, the target feature information including the target detection frame, position, appearance and motion state is extracted, and the detected target feature information is output as the detection result.

[0036] Step 2: Call the MOSSE algorithm for target tracking

[0037] Enter the tracking mode and use the detection results output by the YOLO algorithm to initialize the filter to ensure continuous tracking of the target. The filter tracks the target based on the MOSSE algorithm and outputs tracking results including the target tracking box, position, appearance and motion status.

[0038] Step 3: Switch between detection mode and tracking mode through reward mechanism

[0039] However, the stability of the MOSSE algorithm is easily disturbed during the tracking process, resulting in the problem of target tracking frame drift. Once the target moves out of the field of view due to occlusion, rapid movement or other factors, the MOSSE algorithm will no longer be able to effectively track the target and mark the target status as "lost". By introducing a reward mechanism based on reinforcement learning to dynamically adjust the calling frequency of the YOLO detection algorithm, the MOSSE tracking frame is reset to improve stability.

[0040] Based on the reward mechanism of reinforcement learning, the difference between the detection result output by the YOLO algorithm and the tracking result output by the MOSSE algorithm is used as the reward judgment standard of reinforcement learning, so that reinforcement learning can effectively detect the drift during MOSSE tracking. When the difference between the center point of the detection frame output by the YOLO algorithm and the center point of the tracking frame output by the MOSSE algorithm is greater than or equal to the set threshold, the tracking frame of the MOSSE algorithm drifts, a negative reward is generated, the detection mode is switched, the YOLO algorithm is called to re-detect the target, the latest target features are obtained and passed to the MOSSE algorithm, and the MOSSE tracking frame is updated. When the difference between the center point of the detection frame output by the YOLO algorithm and the center point of the tracking frame output by the MOSSE algorithm is less than the set threshold, a positive reward is generated, the tracking mode is switched to the tracking mode, the MOSSE algorithm is called to track the target, and the target position is updated to ensure the continuity and stability of target tracking.

[0041] When the target is continuously visible, the MOSSE algorithm keeps tracking and updates the target position in real time. When the target is lost again, it returns to the target loss detection stage and calls the YOLO algorithm again for detection.

[0042] Figure 2 It is a flow chart of the YOLO algorithm of the present invention for detecting targets. YOLO uses a deep convolutional neural network for feature extraction. Through multi-layer convolution operations, the model can extract important feature information in the target image, such as edges, shapes, and colors. Compared with the traditional YOLO algorithm, the present invention makes targeted modifications to the labeled and normalized YOLO training data set, uses a satellite model to describe space targets, and constructs target images including diversified data such as different angles and different lighting conditions, so that YOLO can more accurately identify specific objects to be tested, thereby enhancing the detection capability in complex environments.

[0043] Construct a target image including diverse data such as different angles and different lighting conditions, including but not limited to performing enhanced operations such as image rotation, image scaling, and brightness adjustment on the target image.

[0044] Image rotation: For any given target image point (x, y), its new coordinates (x′, y′) at the rotation angle θ are calculated using the rotation matrix, expressed as:

[0045]

[0046] Among them, θ represents the rotation angle, a positive value represents counterclockwise rotation, and a negative value represents clockwise rotation, which can effectively simulate the appearance changes of the target under different viewing angles.

[0047] Image scaling: For each point (x, y) in the target image, its coordinates can be adjusted by the scaling factor s to obtain the new coordinates (x′, y′), expressed as:

[0048]

[0049] If s>1, it means the image is enlarged, and if 0<s<1, it means the image is reduced. This operation can effectively enhance the diversity of training samples.

[0050] Brightness adjustment: To improve the adaptability of the target image under different lighting conditions, contrast and brightness adjustment are used. The original image is set as I, and the adjusted image I′ can be expressed as:

[0051] I′=I·α+β (3)

[0052] Among them, α is the contrast factor, ranging from [0.8 to 2.0]; β is the brightness offset, ranging from [-100 to 100]. This operation can maintain good recognizability under various lighting conditions by adjusting appropriate parameters.

[0053] Diversified data processing can select at least one method to perform at least one-dimensional data enhancement. For example, only image rotation or scaling can be performed to obtain a target image data set of a basic satellite model. Then, two or more data processing methods can be selected to use two-dimensional or multi-dimensional data enhancement methods to further improve the diversity of the target image data set.

[0054] The image target image dataset after the above enhancement operation is trained to obtain the dataset training model, and the dataset training model is embedded in the YOLO algorithm. The YOLO algorithm extracts features of the input space target image according to the dataset training model, which not only improves the accuracy of space target detection, but also enhances the adaptability and robustness of the YOLO algorithm in complex environments. It can effectively cope with various situations such as deformation, illumination change, perspective change, and posture change, ensuring the reliability of detection in dynamic scenes.

[0055] In addition, the YOLO algorithm also introduces a feature pyramid structure, which enables features of different scales to be effectively captured, thereby better handling the detection of multi-scale targets and improving robustness and adaptability in complex environments.

[0056] After feature extraction, YOLO generates predictions at different feature layers, including the bounding box coordinates, class probability, and confidence score of the target. The bounding box is used to locate possible targets, the class probability indicates the possibility of containing a specific target in the box, and the confidence score evaluates the reliability of the prediction.

[0057] The generated bounding boxes and corresponding confidence scores are thresholded to remove predictions below the set threshold. Next, YOLO uses the non-maximum suppression algorithm to further eliminate overlapping bounding boxes, ensuring that only the most likely bounding box is retained for each target, thereby reducing redundant detections and improving the clarity of the detection results. After the above processing, YOLO finally outputs the detected target feature information, including the detection box, position, appearance, and motion state of each target, achieving more efficient and accurate target detection.

[0058] Figure 3 It is a flow chart of the MOSSE algorithm of the present invention for tracking a target. Before starting tracking, the target needs to be initialized first. The target feature information detected by the YOLO algorithm is used as a template, that is, the image portion of the target in the initial frame, and the MOSSE algorithm uses the template to generate a filter to minimize the square error of the output, thereby improving the tracking accuracy. In order to further optimize the efficiency of the MOSSE algorithm, the present invention improves its pre-training process: when initialized for the first time, the number of pre-training iterations is set to 128 times, so as to fully train the filter, make it better adapt to the target, and achieve higher initial performance. Thereafter, the number of pre-training iterations is reduced to 1 to reduce the time of subsequent initialization, so that MOSSE can respond to changes in the target more quickly, significantly speeding up the switching speed between the MOSSE and YOLO algorithms.

[0059] In the single-step iterative tracking process, the MOSSE algorithm processes the current frame target image to find the target. First, the algorithm converts the current frame target image into the frequency domain and applies the previously generated filter for convolution operation to calculate the response value that characterizes the degree of match between each position and the target feature. The higher the response value, the more likely the position contains the target. After convolution, the MOSSE algorithm generates a response map that reflects the degree of match of the target at each position in the current frame image, and finds the point with the largest response value in the response map as the estimated position of the target. After finding the target position, the MOSSE algorithm adaptively updates the filter based on the current observation results of the YOLO algorithm and the current response value of the filter to adapt to the appearance changes of the target.

[0060] In addition, in the existing MOSSE algorithm, the filter update mechanism is usually fixed and will not be dynamically adjusted according to the changes in the external environment or the target. Therefore, the present invention designs an adaptive filter update strategy for the MOSSE algorithm. By dynamically adjusting the weight of the filter according to the changes in the appearance of the target, such as changes in illumination and shape, during the tracking process, the high sensitivity of the tracking algorithm to the changes in the target is ensured. On this basis, the response value threshold is set to 0.65, and when the maximum response value corresponding to the estimated position of the target is lower than the threshold, the filter update is triggered.

[0061] The filter update formula is:

[0062] H t+1 =(1-α)H t +αF t (4)

[0063] Among them, H t+1 is the response value at time t+1, H t is the response value at time t, F t is the target feature information detected by the YOLO algorithm at time t, and α is the learning rate, which determines the weight ratio of new and old information in the filter. The response value threshold is used to determine whether to trigger the filter update, ensuring that when the filter's tracking accuracy for the target decreases, that is, the response value is lower than the threshold, it is adaptively updated in time.

[0064] The adaptive update strategy can dynamically adjust α to make the update process more flexible. When the difference between the maximum response value in the response graph and the maximum response value of the previous frame exceeds the set threshold, it is considered that the target has changed significantly. At this time, α can be increased to between 0.7 and 0.9 to increase the amplitude of the filter update. When the target changes slightly, that is, the difference between the maximum response value and the maximum response value of the previous frame does not exceed the set threshold, α is reduced to between 0.3 and 0.5 to reduce the frequency and amplitude of the update, ensure the stability of the tracking process, and avoid performance degradation caused by frequent updates.

[0065] By introducing the target feature change information provided by the YOLO detection algorithm, when YOLO detects feature changes of the target, it directly triggers the adaptive update of the MOSSE filter, updates the tracking frame to the target features currently detected by YOLO, and then determines the target position by the updated tracking frame, ensuring that the tracker can still accurately track when the appearance changes greatly, thereby improving the robustness of target tracking.

[0066] In order to verify the technical advantage of the present invention in reducing the time delay in the detection and tracking process, the following experimental design was carried out, and the advantages of the method of the present invention were demonstrated through the experimental results.

[0067] (1) Experimental conditions

[0068] A satellite model was selected as the experimental object, and the motion trajectory of the satellite in orbit was simulated. A variety of environmental conditions were set to simulate typical situations in space scenes.

[0069] The experiment compares the traditional MOSSE algorithm with the method of the present invention. For each experimental environment, the time from target loss to re-tracking and the running frame rate during the whole process were recorded. Five tests were conducted under each condition and the average value was taken to ensure the stability of the results.

[0070] (2) Comparison of experimental data

[0071] Table 1

[0072]

[0073] Table 2

[0074]

[0075] (3) Experimental results analysis

[0076] Table 1 shows the tracking time comparison between the traditional MOSSE algorithm and the method of the present invention. It can be seen that the method of the present invention reduces the time of the lost and then tracked process in all test environments, especially in the case of normal light and fast movement, the time is saved by 58.46%; Table 2 shows the running frame rate comparison between the traditional MOSSE algorithm and the method of the present invention. It can be concluded from the table that the method of the present invention significantly improves the frame rate in various environments and improves the real-time performance. Especially in the case of low light, the running frame rates of the target at normal speed and fast movement are increased by 155.56% and 165.62% respectively, which is a significant improvement.

[0077] By optimizing and improving the MOSSE algorithm, the present invention significantly reduces the algorithm startup time and the time to re-track the target after switching. The traditional MOSSE algorithm has a large number of iterations, resulting in a slow startup speed, while the present invention greatly speeds up the startup and switching speeds by shortening the number of iterations. In addition, the target detection and tracking method proposed in the present invention has a significant improvement in the running frame rate. The running frame rate of the traditional MOSSE algorithm is approximately between 7 and 11fps, while the running frame rate of the method of the present invention has been increased to 18 to 22fps, further improving the effect of real-time tracking.

[0078] In summary, the present invention provides a highly robust target loss detection and fast tracking method, which comprises the following steps: when the target is detected to be lost, switching to the YOLO algorithm for real-time detection to confirm whether the target reappears; after successful detection, extracting the target features and passing them to the MOSSE algorithm for fast tracking; if the target is lost again, repeating the above process, automatically returning to the detection stage and switching to the YOLO algorithm.

[0079] Through the flexible switching between detection and tracking, the present invention has excellent adaptability and stability when dealing with complex spatial environments and dynamic changes of targets, realizes fast and efficient target detection and recovery tracking, and improves the robustness of the tracking system.

[0080] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above specific embodiments. The above specific embodiments and the description in the specification are only for further illustrating the principles and effects of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A robust target loss detection and fast tracking method, characterized in that: The steps include: Step 1: Initialize the space target to a lost state, enter the detection mode, call the YOLO algorithm to extract the feature information of the space target, and obtain the target detection result; Step 2, entering the tracking mode, initializing the MOSSE algorithm according to the target detection result, pre-training the filter generated by the initialization MOSSE algorithm, and obtaining the target tracking result; Step three: learning the difference between the target detection result and the target tracking result through a reward mechanism, and switching between the detection mode and the tracking mode according to the learning result of the reward mechanism.

2. A highly robust target loss detection and fast tracking method according to claim 1, characterized in that: The step 1 calls the YOLO algorithm to extract spatial target feature information, which specifically includes the following steps: Step 1-1, annotating and normalizing the space target image, using a satellite model to describe the space target, constructing a space target image including diversified data of different angles and different lighting conditions, training the space target image of the diversified data, and obtaining a space target image data set training model; Step 1-2, extract features from the input space target image based on the model trained by the space target image dataset.

3. A highly robust target loss detection and fast tracking method according to claim 1, characterized in that: The second step is to pre-train and initialize the filter generated by the MOSSE algorithm. When the filter is initialized for the first time, the number of pre-training iterations is set to the maximum value. When the filter is initialized subsequently, the number of pre-training iterations is set to 1. The process of pre-training single-step iteration specifically includes the following steps: Step 2-1, convert the current frame target image into the frequency domain, and apply the filter generated by the initialized MOSSE algorithm to perform convolution operation, calculate the response value that characterizes the degree of matching between each position and the target feature, obtain the current frame response map, and take the position corresponding to the maximum response value as the estimated position of the spatial target; Step 2-2: when the maximum response value is lower than the set threshold, proceed to step 2-3; otherwise, return to step 2-1; Step 2-3, based on the current response value of the filter and the current target detection result obtained by the YOLO algorithm, the filter is adaptively updated, the target position is updated, and the target tracking result is output.

4. The method for detecting and rapidly tracking a target with high robustness according to claim 1, characterized in that: The step three is to learn the difference between the target detection result and the target tracking result through a reward mechanism based on a reinforcement learning algorithm: when the difference between the target detection result and the target tracking result is less than a set threshold, output a positive reward for switching to the tracking mode; When the difference between the target detection result and the target tracking result is greater than or equal to the set threshold, the output switches to the negative reward of the detection mode.

5. The method for detecting and rapidly tracking a target with high robustness according to claim 2, characterized in that: The specific method of step 1-1 to construct a space target image including diversified data of different angles and different lighting conditions is: performing enhancement operations including but not limited to image rotation, image scaling, and brightness adjustment on the space target image.

6. The method for detecting and rapidly tracking a target with high robustness according to claim 2, characterized in that: The step 1-2 extracts features from the input space target image based on the space target image data set training model, specifically: obtaining feature layers of different scales through a feature pyramid structure, generating predictions including bounding boxes, category probabilities, and confidence scores on feature layers of different scales, removing bounding boxes and predictions with confidence scores lower than a set threshold, eliminating overlapping bounding boxes through a non-maximum suppression algorithm, and outputting target detection results.

7. A highly robust target loss detection and fast tracking method according to claim 5, characterized in that: The image rotation rotates the given space target image point (x, y) by an angle θ to obtain the rotated space target image point (x′, y′). A positive value of θ indicates counterclockwise rotation, and a negative value of θ indicates clockwise rotation.

8. A highly robust target loss detection and fast tracking method according to claim 5, characterized in that: The image scaling adjusts a given space target image point (x, y) by a scaling factor s to obtain a scaled space target image point (x′, y′). s>1 means the image is enlarged, and 0<s<1 means the image is reduced.

9. A highly robust target loss detection and fast tracking method according to claim 5, characterized in that: The brightness adjustment performs contrast and brightness adjustment on the original space target image I to obtain an adjusted space target image I′, where I′=I·α+β, α is a contrast factor, and the value range of α is [0.8, 2.0], and β is a brightness offset, and the value range of β is [-100, 100].

10. The method for detecting and rapidly tracking a target with high robustness according to claim 3, characterized in that: In step 2-3, the expression of adaptively updating the filter is based on the current response value of the filter and the current target detection result obtained by the YOLO algorithm: t+1 =(1-α)H t +αF t , where H t+1 is the response value at time t+1, H t is the response value at time t, F t is the target feature information detected by the YOLO algorithm at time t, α is the learning rate, when the difference between the maximum response value in the current frame response map and the maximum response value in the previous frame response map is greater than the set threshold, the value of α is 0.7-0.9, when the difference between the maximum response value in the current frame response map and the maximum response value in the previous frame response map is less than or equal to the set threshold, the value of α is 0.3-0.5.

Images