A UAV target tracking method based on embodied intelligence

Abstract

This invention discloses a drone target tracking method based on embodied intelligence. By introducing reinforcement learning based on an improved YOLOv5 target recognition algorithm, combined with multi-level actions, drone motion is controlled through weighted iteration. This allows the drone to upgrade its actions under specific conditions, more efficiently circumvent obstacles and achieve a better view of the target, thereby significantly improving the drone's target tracking success rate. Compared to traditional target tracking methods, this method has a higher target tracking success rate and can effectively complete drone target tracking tasks, meeting the target tracking needs of drones in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of drone target tracking, and in particular to a drone target tracking method based on embodied intelligence. Background Art

[0002] The low-altitude economy plays an increasingly important role in promoting economic development, strengthening social security, and serving national defense. The drone industry, as a key component of the low-altitude economy, is also experiencing rapid growth. Therefore, intelligent target tracking for drones, as a key technology for drones to perform various missions, holds significant economic and strategic significance. In recent years, the development of drone target tracking technology has become a key application in multiple fields. In particular, its application has become increasingly widespread and important in military reconnaissance, disaster monitoring, and agricultural surveying. In the military, drone target tracking enables real-time monitoring and tracking of enemy targets, providing crucial intelligence support for operational decision-making. In disaster monitoring, drones can quickly and accurately detect and track signs of life within disaster-stricken areas, enabling rescue workers to rapidly deploy relief operations. In agricultural surveying, drones can monitor pests and diseases in farmland, reduce pesticide use through precise positioning and application, and improve the efficiency and quality of agricultural production. However, as mission requirements diversify and become increasingly complex, traditional target tracking methods have a low tracking success rate, making them inadequate for drone target tracking in complex environments.

[0003] Alhafnawi M, Salameh HB, et al. reviewed traditional UAV target tracking methods in their paper "A Survey of Indoor and Outdoor UAV-Based Target Tracking Systems: Current Status, Challenges, Technologies, and Future Directions." Traditional target tracking methods are often based on image processing and pattern recognition technologies, such as feature point matching or motion prediction. However, these methods have some defects when faced with complex and changing environments. In dense crowds or complex backgrounds, traditional feature point matching methods may fail, resulting in target loss or misjudgment. In addition, because traditional methods often rely on the processing of static images, they are more sensitive to changes in the target's shape and posture, making it difficult to accurately track dynamic targets. Therefore, traditional target tracking algorithms have certain limitations.

[0004] On the other hand, Azar AT, Koubaa A, et al. reviewed reinforcement learning-based drone target tracking methods in their paper "Drone Deep Reinforcement Learning: A Review." Reinforcement learning, as a trial-and-error learning paradigm, optimizes tracking performance by allowing drones to continuously adjust their strategies in practice. However, this approach also presents some challenges. First, reinforcement learning requires a large amount of training data and time to converge to a better strategy, and in practical applications, this training cost is often not negligible. Second, because drone target tracking often occurs in complex and changing environments, reinforcement learning algorithms may be affected by environmental noise and interference, causing instability and poor tracking results. Therefore, while reinforcement learning-based methods have achieved certain results in some scenarios, their adaptability to the environment still presents significant challenges. Summary of the Invention

[0005] The purpose of this invention is to disclose a drone target tracking method based on embodied intelligence. This method introduces reinforcement learning based on an improved YOLOv5 target recognition algorithm, combines multi-level actions, and controls drone motion through weight iteration. This allows the drone to upgrade its own actions under specific conditions, more efficiently bypass obstacles to obtain a better target field of view, and thus significantly improve the success rate of drone target tracking. Compared with traditional target tracking methods, this method has a higher target tracking success rate and can effectively complete drone target tracking tasks, meeting the target tracking task requirements of drones in complex environments (such as scenarios where the target is partially obscured).

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A method for tracking a target using an unmanned aerial vehicle (UAV) based on embodied intelligence, comprising the following steps:

[0008] S1, based on the YOLOv5 framework, adds the Biformer attention mechanism to the backbone module, retains the original YOLOv5s detection head structure, adds a detection head structure, and then selects Focal EIoU Loss as the loss function to establish an object recognition model;

[0009] S2, collects small target datasets and manually annotates them to generate training datasets for training and evaluating target recognition models;

[0010] S3, initialize the primary action weight and secondary action weight of the drone, where the primary action weight includes the weights of the four movement directions: forward, backward, left, and right, and the secondary action weight includes the weight of clockwise movement and the weight of counterclockwise movement;

[0011] S4, input k consecutive frames of field of view images taken by the onboard camera into the target recognition model to obtain k target confidences and calculate the initial average confidence;

[0012] S5, based on the first-level action weight, controls the first-level motion direction of the drone at the next moment according to probability. Given the first-level action random variable X and the four values ​​corresponding to the first-level action, calculate the probability of the drone's action at the next moment; in the new state after the drone completes the first-level action X, calculate the average confidence in the new state, and combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's first-level action X. X1 ; Add the feedback of the first-level action X to the weight of the first-level action X, update the weight of the first-level action in the corresponding direction, and normalize the weight of the first-level action;

[0013] S6, repeat step S5 until t times e X1 <0, go to step S7;

[0014] S7, based on the secondary action weight, controls the drone's motion direction at the next moment according to probability. Given the secondary action random variable X* and the two values ​​corresponding to the secondary action, calculate the drone's action probability at the next moment; in the new state after the drone's secondary action X* is completed, calculate the average confidence in the new state. Combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's secondary action X*. X2 ; Add the feedback of the secondary action X* to the weight of the secondary action X*, update the secondary action weight in the corresponding direction, and normalize the secondary action weight;

[0015] S8, repeat step S7 until t times e X2 <0;

[0016] S9, repeat steps S5 to S8 until 2t times in a row X1 and e X2 If both are less than 0, the tracking task is completed.

[0017] As a preferred example, step S1 further includes:

[0018] The BiFormer attention mechanism is introduced on the backbone of YOLOv5. Specifically, the BiFormer, a universal visual transformer based on the BRA module, adopts a four-level pyramid structure to focus on a small number of relevant tags in a query-adaptive manner. Overlapping block embedding is used in the first stage, and block merging modules are used in the second to fourth stages to reduce the input spatial resolution while increasing the number of channels. Successive BiFormer blocks are then used for feature transformation. At the beginning of each block, a 3*3 depthwise convolution is used to implicitly encode relative position information. Subsequently, the BRA module and a two-layer multi-layer perceptron module with an expansion rate of e are applied in sequence for cross-position relationship modeling and per-position embedding, respectively.

[0019] The neck network uses a bidirectional cross-connection approach to perform feature fusion, obtaining a series of hierarchical feature representations, which are then passed to the detection head. The original YOLOv5s detection head structure is retained, and a convolutional layer is added to obtain a 160x160 feature map. Feature fusion is then performed using information retained in the low-level high-dimensional feature distribution space and information from the deep low-dimensional feature space.

[0020] Select Focal EIoU Loss as the loss function; given the prediction box B and the object box B gt Define EIoU Loss as follows:

[0021]

[0022] Where IOU refers to the intersection area of ​​the target box predicted by the model and the real target box divided by the area of ​​the union of the two, h gt Refers to the height of the real target box, w gt refers to the width of the real target box, h refers to the height of the target box predicted by the model, and w refers to the width of the target box predicted by the model; b and b gt Indicates B and B gt The center point of the prediction box, ρ(·) represents the Euclidean distance, c is the diagonal length of the smallest enclosing box that can contain the prediction box and the object box, and the variable h c and w c Represents the height and width of the minimum bounding box respectively, and then introduces Focal Loss to use (1-IOU) γ The weight of the sample is used to adjust the sample weight, balancing the small loss of samples with good regression quality and the large loss of samples with poor regression quality; Focal EIoU Loss is defined as:

[0023] L focal-eiou =-(1-IOU) γ log(IOU)L eiou .

[0024] As a preferred example, step S2 further includes:

[0025] In different scenarios, weather, and lighting conditions, at an altitude of 4060 meters above the ground, different types of drone platforms were used to collect target tracking datasets; the datasets were manually annotated; the training data was enhanced using Mosaic or Mixup; and the target recognition model was trained and evaluated.

[0026] As a preferred example, in step S3, the first-level action weight of the drone is initialized as follows:

[0027] W forward，0 =W back，0 =W left，0 =W right，0 =0.25

[0028] Where W forward，0 、W back，0 、W left，0 、W right，0 are the weights of the four movement directions: forward, backward, left and right;

[0029] Initialize the drone's secondary action weights as follows:

[0030] W cw，0 =W ccw，0 =0.5

[0031] Where W cw，0 and W ccw，0 They are the clockwise and counterclockwise movement weights respectively.

[0032] As a preferred example, step S4 further includes:

[0033] Input k consecutive frames of field of view pictures taken by the onboard camera into the target recognition model to obtain k target confidences c1, c2, ..., c k , calculate the initial average confidence c0:

[0034]

[0035] As a preferred example, step S4 further includes:

[0036] According to the weight of the first-level action, the movement direction of the drone at the next moment is controlled according to probability. Given the first-level action random variable X, and the four values ​​X corresponding to the first-level action forward , X back , X left , X right , the calculated action probability of the drone at the next moment is as follows:

[0037]

[0038] In the new state after the drone action is completed, the average confidence in the new state is calculated Calculate the state update feedback due to action X

[0039] Add the feedback of the first-level action X to the weight of the first-level action X and update the weight of the first-level action in the corresponding direction.

[0040] As a preferred example, step S4 further includes:

[0041] Normalize the first-level action weights:

[0042]

[0043] Where W forward ′、W back ′、W left ′ and W right ′ represents the action weights of the first-level action in the new state in the four directions of forward, backward, left and right; W forward 、W back 、W left 、W right They are the weights of the four movement directions forward, backward, left and right after the last round of weight update.

[0044] As a preferred example, step S7 further includes:

[0045] According to the secondary action weight, the next moment's motion direction of the drone is controlled by probability. Given the secondary action random variable X* and the two values ​​X corresponding to the secondary action cw , X ccw , the calculated action probability of the drone at the next moment is as follows:

[0046]

[0047] In the new state after the drone action is completed, the average confidence in the new state is calculated This results in the calculation of the state update feedback caused by action X*

[0048] Add the feedback of action X* to the weight of action X* and update the weight of the secondary action in the corresponding direction.

[0049] As a preferred example, step S7 further includes:

[0050] Normalize the secondary action weights:

[0051]

[0052] Among them, W cw and W ccw are the clockwise and counterclockwise motion weights after the last round of weight update, W cw ′ and W ccw ′ respectively represent the clockwise and counterclockwise action weights of the secondary action in the new state.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] First, the UAV target tracking method based on embodied intelligence of the present invention establishes an efficient target recognition model, adds a BiFormer attention module on the basis of the YOLOv5 backbone, and adds an additional high-resolution detection head in the detection layer to improve the backbone network's feature extraction ability for small targets. At the same time, the Focal EIoU Loss is used to optimize the loss function definition, which solves the problem of imbalance between difficult and easy samples, optimizes the network training effect, and improves the accuracy of UAV small target recognition.

[0055] Second, the drone target tracking method based on embodied intelligence of the present invention introduces reinforcement learning, combines multi-level actions, and controls the drone movement through weight iteration, so that the drone can upgrade its own actions under specific conditions, more efficiently bypass obstacles to obtain a better target view, and significantly improve the success rate of drone target tracking.

[0056] Third, the embodied intelligence-based drone target tracking method of this invention offers enhanced real-time performance and stability. By optimizing the target recognition framework and incorporating embodied intelligence into the target tracking algorithm, drones can more quickly and accurately identify and track targets. This efficient target recognition and tracking system can help drones better cope with various challenges in complex environments, maintain stable operation, and respond promptly, better meeting practical application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the UAV target tracking method based on embodied intelligence proposed in the present invention;

[0058] Figure 2 This is a framework diagram of the target recognition model based on the improved YOLOv5 proposed in the present invention;

[0059] Figure 3 This is a principle block diagram of the embodied intelligence algorithm adopted by the present invention. DETAILED DESCRIPTION

[0060] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0061] The present invention discloses a drone target tracking method based on embodied intelligence, the drone target tracking method comprising the following steps:

[0062] S1, based on the YOLOv5 framework, adds the Biformer attention mechanism to the backbone module, retains the original YOLOv5s detection head structure, adds a detection head structure, and then selects Focal EIoU Loss as the loss function to establish an object recognition model;

[0063] S2, collects small target datasets and manually annotates them to generate training datasets for training and evaluating target recognition models;

[0064] S3, initialize the primary action weight and secondary action weight of the drone, where the primary action weight includes the weights of the four movement directions: forward, backward, left, and right, and the secondary action weight includes the weight of clockwise movement and the weight of counterclockwise movement;

[0065] S4, input k consecutive frames of field of view images taken by the onboard camera into the target recognition model to obtain k target confidences and calculate the initial average confidence;

[0066] S5, based on the first-level action weight, controls the first-level motion direction of the drone at the next moment according to probability. Given the first-level action random variable X and the four values ​​corresponding to the first-level action, calculate the probability of the drone's action at the next moment; in the new state after the drone completes the first-level action X, calculate the average confidence in the new state, and combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's first-level action X. X1 ; Add the feedback of the first-level action X to the weight of the first-level action X, update the weight of the first-level action in the corresponding direction, and normalize the weight of the first-level action;

[0067] S6, repeat step S5 until t times e X1 <0, go to step S7;

[0068] S7, based on the secondary action weight, controls the drone's motion direction at the next moment according to probability. Given the secondary action random variable X* and the two values ​​corresponding to the secondary action, calculate the drone's action probability at the next moment; in the new state after the drone's secondary action X* is completed, calculate the average confidence in the new state. Combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's secondary action X*. X2; Add the feedback of the secondary action X* to the weight of the secondary action X*, update the secondary action weight in the corresponding direction, and normalize the secondary action weight;

[0069] S8, repeat step S7 until t times e X2 <0;

[0070] S9, repeat steps S5 to S8 until 2t times in a row X1 and e X2 If both are less than 0, the tracking task is completed.

[0071] Compared to traditional target tracking methods and those based on reinforcement learning, embodied intelligence (Embodied Intelligence) offers a radically new concept. It emphasizes achieving intelligent behavior through actual interaction with the environment, rather than relying solely on analysis and learning from static data. In drone target tracking, embodied intelligence allows the drone to continuously accumulate experience through actual flight and target tracking, thereby gradually improving its tracking effectiveness. Compared to traditional methods, embodied intelligence places greater emphasis on real-time interaction between the drone and the environment, enabling it to better adapt to complex and changing tracking environments and possessing greater adaptability and generalization capabilities.

[0072] Based on the above analysis, embodied intelligence is a proven and effective solution for drone target tracking. By enabling drones to interact and learn from their environment in real time, continuously improving their tracking strategies, they can enhance their adaptability and tracking effectiveness in complex environments, thereby better meeting the needs of practical applications. The introduction of embodied intelligence not only overcomes the limitations of traditional methods in complex environments but also overcomes the reliance of reinforcement learning methods on training data and environmental noise, providing new insights and approaches for the further development of drone target tracking technology.

[0073] like Figure 1 As shown, the present invention proposes a drone target tracking method based on embodied intelligence, which includes:

[0074] Step 1: Establish an efficient target recognition framework.

[0075] Based on the YOLOv5 framework, the optimized target recognition framework is shown in the attached figure. Figure 2 shown.

[0076] The BiFormer attention mechanism is introduced on the YOLOv5 backbone, reducing computational effort while preserving fine-grained image detail and improving the backbone network's ability to extract features for small objects. The BiFormer, a universal visual transformer based on the BRA module, employs a four-level pyramid structure to focus on a small number of relevant markers in a query-adaptive manner, using overlapping block embedding in the first stage. In the second to fourth stages, a block merging module is used to reduce the input spatial resolution while increasing the number of channels, followed by feature transformation using successive BiFormer blocks. A 3x3 depthwise convolution is used at the beginning of each block to implicitly encode relative position information. Subsequently, the BRA module and a two-layer multilayer perceptron module with an expansion rate of e are applied sequentially for cross-position relationship modeling and per-position embedding, respectively.

[0077] The neck network uses bidirectional cross-connections for feature fusion, generating a series of hierarchical feature representations that are then passed to the detection head. The original YOLOv5s detection head structure is retained, with one convolutional layer added to produce a 160x160 feature map. This large-scale feature map effectively improves the accuracy of small object detection. Feature fusion is then performed by combining information retained in the lower-level, high-dimensional feature distribution space with information in the deeper, low-dimensional feature space.

[0078] Select Focal EIoU Loss as the loss function. Given the prediction box B and the object box B gt Define EIoU Loss as follows:

[0079]

[0080] where b and b gt Indicates B and B gt The center point of , ρ(·) represents the Euclidean distance, for example ρ(b, b gt )=||bb gt ||2, c is the diagonal length of the minimum enclosing box that can contain the prediction box and the object box, variable h c and w c Represents the height and width of the minimum bounding box respectively, and then introduces Focal Loss to use (1-IOU) γ The weight of the sample is used to adjust the sample weight, balancing the small loss of samples with good regression quality and the large loss of samples with poor regression quality. The Focal EIoU Loss is defined as:

[0081] L focal-eiou =-(1-IOU) γ log(IOU)L eiou .

[0082] Step 2: Recognition model training.

[0083] A target dataset was collected and manually annotated using different drone platforms in various scenarios, weather, and lighting conditions, at an altitude of 40-60 meters above the ground. This dataset reflects the general use of drones in real-world applications and is consistent with the background of this invention. During network training, the input training data was augmented using methods such as Mosaic and Mixup. The images were then adaptively scaled and anchor boxes were calculated. Finally, the framework proposed in step 1 was trained to derive a weighted model.

[0084] Step 3: Initialize the drone action weights.

[0085] The embodied intelligence algorithm block diagram is as follows Figure 3 shown.

[0086] Initialize the weights of the drone's first-level actions forward, backward, left, and right as follows:

[0087] W forward，0 =W back，0 =W left，0 =W right，0 =0.25

[0088] Where W forward，0 、W back，0 、W left，0 、W right，0 are the weights of the four movement directions: forward, backward, left and right;

[0089] Initialize the drone's secondary action weights as follows:

[0090] W cw，0 =W ccw，0 =0.5

[0091] Where W cw，0 and W ccw，0 They are the clockwise and counterclockwise movement weights respectively.

[0092] Step 4: Calculate the average confidence.

[0093] Input k consecutive frames of field of view images taken by the onboard camera into the recognition framework in step 1 to obtain k target confidences, denoted as c1, c2, ..., c k .

[0094] Calculate the average confidence as follows:

[0095]

[0096] Step 5: Calculate the action probability.

[0097] According to the weight of the first-level action, the movement direction of the drone at the next moment is controlled according to probability. Given the action random variable X, and the four values ​​X corresponding to the first-level action forward 、X back 、X left 、X right , the probability of the drone’s action at the next moment is as follows:

[0098]

[0099] Step 6: Calculate action feedback.

[0100] In the new state after the drone action is completed, repeat step 4 to obtain the average confidence in this state Calculate the state update feedback due to action X

[0101] Step 7: Update the first-level action weights.

[0102] Add the feedback of action X to the weight of action X to make the forward action X = X forward For example, get feedback e forward Then update the weight W forward W forward +e forward .

[0103] Normalize the first-level action weights as follows:

[0104]

[0105] Where W forward ′、W back ′、W left ′、W right ′ represents the weights of the forward, backward, left, and right actions in the new state respectively.

[0106] Step 8, repeat steps 5 to 7 until t times X1 <0, go to step 9.

[0107] Step 9: Enable the secondary action and update the secondary action weight.

[0108] According to the secondary action weight, the next moment's motion direction of the drone is controlled by probability. Given the action random variable X*, and the two values ​​X corresponding to the secondary action cw , X ccw , the probability of the drone’s action at the next moment is as follows:

[0109]

[0110] In the new state after the drone action is completed, repeat step 4 to obtain the average confidence in this state This results in the calculation of the state update feedback caused by action X* Add the feedback of action X* to the weight of action X* to make the clockwise action X*=X cw For example, get feedback e cw Then update the weight W cw W cw +e cw .

[0111] Normalize the secondary action weights as follows:

[0112]

[0113] Where W cw ′ and W ccw ′ respectively represent the clockwise and counterclockwise action weights of the secondary action in the new state.

[0114] Step 10, repeat step 9 until t times in a row X2 <0, return to step 8.

[0115] Step 11, repeat step 10 until 2t times in a row X1 <0 and e X2 <0, the tracking task ends.

[0116] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions for executing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0120] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0121] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A drone target tracking method based on embodied intelligence, characterized in that: The UAV target tracking method comprises the following steps: S1, based on the YOLOv5 framework, adds the Biformer attention mechanism to the backbone module, retains the original YOLOv5s detection head structure, adds a detection head structure, and then selects Focal EIoU Loss as the loss function to establish an object recognition model; S2, collects small target datasets and manually annotates them to generate training datasets for training and evaluating target recognition models; S3, initialize the primary action weight and secondary action weight of the drone, where the primary action weight includes the weights of the four movement directions: forward, backward, left, and right, and the secondary action weight includes the weight of clockwise movement and the weight of counterclockwise movement; S4, input k consecutive frames of field of view images taken by the onboard camera into the target recognition model to obtain k target confidences and calculate the initial average confidence; S5, based on the first-level action weight, controls the first-level motion direction of the drone at the next moment according to probability. Given the first-level action random variable X and the four values ​​corresponding to the first-level action, calculate the probability of the drone's action at the next moment; in the new state after the drone completes the first-level action X, calculate the average confidence in the new state, and combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's first-level action X. X1 ; Add the feedback of the first-level action X to the weight of the first-level action X, update the weight of the first-level action in the corresponding direction, and normalize the weight of the first-level action; S6, repeat step S5 until t times e X1 <0, go to step S7; S7, based on the secondary action weight, controls the drone's motion direction at the next moment according to probability. Given the secondary action random variable X* and the two values ​​corresponding to the secondary action, calculate the drone's action probability at the next moment; in the new state after the drone's secondary action X* is completed, calculate the average confidence in the new state. Combine the average confidence in the new state with the initial average confidence to calculate the state update feedback e caused by the drone's secondary action X*. X2 ; Add the feedback of the secondary action X* to the weight of the secondary action X*, update the secondary action weight in the corresponding direction, and normalize the secondary action weight; S8, repeat step S7 until t times e X2 <0; S9, repeat steps S5 to S8 until 2t times in a row X1 and e X2 If both are less than 0, the tracking task is completed.

2. The method for tracking drone targets based on embodied intelligence according to claim 1, characterized in that: Step S1 further comprises: The BiFormer attention mechanism is introduced on the backbone of YOLOv5. Specifically, the BiFormer, a universal visual transformer based on the BRA module, adopts a four-level pyramid structure to focus on a small number of relevant tags in a query-adaptive manner. Overlapping block embedding is used in the first stage, and block merging modules are used in the second to fourth stages to reduce the input spatial resolution while increasing the number of channels. Successive BiFormer blocks are then used for feature transformation. At the beginning of each block, a 3*3 depthwise convolution is used to implicitly encode relative position information. Subsequently, the BRA module and a two-layer multi-layer perceptron module with an expansion rate of e are applied in sequence for cross-position relationship modeling and per-position embedding, respectively. The neck network uses a bidirectional cross-connection approach to perform feature fusion, obtaining a series of hierarchical feature representations, which are then passed to the detection head. The original YOLOv5s detection head structure is retained, and a convolutional layer is added to obtain a 160×160 feature map. Feature fusion is then performed using information retained in the low-level high-dimensional feature distribution space and information from the deep low-dimensional feature space. Select Focal EIoU Loss as the loss function; given the prediction box B and the object box B gt Define EIoU Loss as follows: Where IOU refers to the intersection area of ​​the target box predicted by the model and the real target box divided by the area of ​​the union of the two, h gt Refers to the height of the real target box, w gt refers to the width of the real target box, h refers to the height of the target box predicted by the model, and w refers to the width of the target box predicted by the model; b and b gt Indicates B and B gt The center point of the prediction box, ρ(·) represents the Euclidean distance, c is the diagonal length of the smallest enclosing box that can contain the prediction box and the object box, and the variable h c and w c Represents the height and width of the minimum bounding box respectively, and then introduces Focal Loss to use (1-IOU) γ The weight of the sample is used to adjust the sample weight, balancing the small loss of samples with good regression quality and the large loss of samples with poor regression quality; Focal EIoU Loss is defined as: L focal-eiou =-(1-IOU) γ log(IOU)L eiou 。 3. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S2 further comprises: In different scenarios, weather, and lighting conditions, at an altitude of 40-60 meters above the ground, different types of drone platforms were used to collect target tracking datasets; the datasets were manually annotated; the training data was enhanced using Mosaic or Mixup; and the target recognition model was trained and evaluated.

4. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: In step S3, the first-level action weights of the drone are initialized as follows: IN forward，0 =In back，0 =In left，0 =In right，0 =0.25 Where W forward，0 、W back，0 、W left，0 、W right，0 are the weights of the four movement directions: forward, backward, left and right; Initialize the drone's secondary action weights as follows: IN cw，0 =In ccw，0 =0.5 Where W cw，0 and W ccw，0 They are the clockwise and counterclockwise movement weights respectively.

5. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S4 further comprises: Input k consecutive frames of field of view pictures taken by the onboard camera into the target recognition model to obtain k target confidences c1, c2, ..., c k , calculate the initial average confidence 6. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S4 further comprises: According to the weight of the first-level action, the movement direction of the drone at the next moment is controlled according to probability. Given the first-level action random variable X, and the four values ​​X corresponding to the first-level action forward , X back , X left , X right , the calculated action probability of the drone at the next moment is as follows: In the new state after the drone action is completed, the average confidence in the new state is calculated Calculate the state update feedback due to action X Add the feedback of the first-level action X to the weight of the first-level action X and update the weight of the first-level action in the corresponding direction.

7. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S4 further comprises: Normalize the first-level action weights: Where W forward ′、W back ′、W left ′ and W right ′ represents the action weights of the first-level action in the new state in the four directions of forward, backward, left and right; W forward 、W back 、W left 、W right They are the weights of the four movement directions forward, backward, left and right after the last round of weight update.

8. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S7 further includes: According to the secondary action weight, the next moment's motion direction of the drone is controlled by probability. Given the secondary action random variable X* and the two values ​​X corresponding to the secondary action cw , X ccw , the calculated action probability of the drone at the next moment is as follows: In the new state after the drone action is completed, the average confidence in the new state is calculated This results in the calculation of the state update feedback caused by action X* Add the feedback of action X* to the weight of action X* and update the weight of the secondary action in the corresponding direction.

9. The method for tracking drone targets based on embodied intelligence according to claim 1, wherein: Step S7 further includes: Normalize the secondary action weights: Among them, W cw and W ccw are the clockwise and counterclockwise motion weights after the last round of weight update, W cw ′ and W ccw ′ respectively represent the clockwise and counterclockwise action weights of the secondary action in the new state.

Images