A deep learning-based unmanned aerial vehicle target dynamic tracking method

Abstract

The application discloses a kind of unmanned aerial vehicle target dynamic tracking methods based on deep learning, including the following steps: continuous image sequence, unmanned aerial vehicle attitude information and camera parameter are collected in the process of unmanned aerial vehicle flight, generate observation sequence and platform state sequence;Target initialization is carried out, and target tracking state is generated;State representation inference is carried out, and multiple target state reversibility representations are generated;Multiple target state reversibility representations are accumulated, and target state reversibility constraint domain is generated;Current target state reversibility representation is mapped to target state reversibility constraint domain, and reversibility verification result is generated;Target tracking task execution certificate is generated based on reversibility verification result;State disposal is carried out based on target tracking task execution certificate.The application is based on the target state reversibility modeling method of deep learning, realizes unmanned aerial vehicle target dynamic stable tracking, with the advantages of high tracking precision, strong stability and excellent abnormal suppression capability.

Description

Technical Field

[0001] This invention relates to the field of dynamic target tracking for unmanned aerial vehicles (UAVs), and more particularly to a deep learning-based method for dynamic target tracking of UAVs. Background Technology

[0002] With the rapid development of drone technology, target dynamic tracking has been widely used in aerial photography, surveillance, autonomous driving and other fields. Current target tracking methods mostly use traditional computer vision and filtering algorithms, such as Kalman filtering and mean drift algorithm. These methods rely on simple motion models and observation data and are suitable for simple and ideal environments. However, in complex flight environments, especially when the attitude and observation angle of the drone change rapidly, existing technologies are unable to cope with the complex changes in target dynamics, resulting in decreased tracking accuracy or even loss of target.

[0003] Existing technologies have significant shortcomings in handling target state updates, predictions, and task execution control. In particular, when the target is moving rapidly or the environment changes significantly, traditional methods cannot effectively determine whether the target tracking task still has the conditions to continue. They usually force the output of tracking results, resulting in mistracking, state accumulation errors, or tracking distortion. Therefore, how to accurately and stably track targets in complex environments and intelligently adjust task execution according to changes in the target state has become an urgent technical problem to be solved. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based method for dynamic tracking of UAV targets. This invention uses a deep learning-based target state reversibility modeling method to achieve dynamic and stable tracking of UAV targets, which has the advantages of high tracking accuracy, strong stability and excellent anomaly suppression capability.

[0005] A deep learning-based method for dynamic target tracking of unmanned aerial vehicles (UAVs) according to an embodiment of the present invention includes the following steps: During the flight of the UAV, a continuous image sequence is acquired and arranged according to the time index to generate an observation sequence. At the same time, the UAV attitude information and camera parameters are acquired and aligned according to the same time index to generate a platform status sequence. Based on the observation sequence and platform state sequence, a target candidate set is generated, the target is initialized, and a target tracking state is generated. The target tracking state includes a target identity identifier, a target state vector, and a target historical state buffer. Based on the target tracking state, the target state vector sequence is obtained from the target historical state buffer. The target state vector sequence, observation sequence and platform state sequence are input into the deep learning model to perform state representation inference and generate multiple target state invertibility representations corresponding to different target state vectors. Accumulate multiple target state reversibility representations to generate a target state reversibility constraint domain; Map the current target state reversibility representation to the target state reversibility constraint domain to generate reversibility verification results; Based on the reversibility verification results, a target tracking task execution certificate corresponding to the target identity is generated. The target tracking task execution certificate includes a valid status, a downgraded status, and a revoked status. The target tracking task is state-managed based on the target tracking task execution certificate. When it is in a valid state, the target state vector is updated and written to the target historical state buffer. When it is in a degraded state, the target historical state buffer is not updated. When it is in a revoked state, the target tracking task is terminated.

[0006] Optionally, the generation of the target tracking state specifically includes: Under each time index, target candidate generation processing is performed on the observation sequence, target regions in the observation sequence are detected and located, target regions that meet the detection conditions are extracted, and an initial target candidate set is generated. The initial target candidate set is associated with the platform state sequence. Based on the UAV attitude information and camera parameters in the platform state sequence, spatial consistency screening is performed on each target candidate in the initial target candidate set to remove target candidates that do not meet the spatial constraints and generate a filtered target candidate set. For each target candidate in the filtered target candidate set, target initialization processing is performed to assign a corresponding target identity identifier to each target candidate; Based on the spatial location of the target candidate in the observation sequence and the UAV attitude information and camera parameters in the platform state sequence, a corresponding target state vector is generated for each target identity. Using the target identity as an index, a corresponding target history state buffer is established for each target identity. The target state vectors generated under each time index are written into the corresponding target history state buffers in order of time index. The target identity identifier, target state vector, and target historical state buffer are bound together to generate target tracking state. When there are multiple target candidates in the target candidate set, a corresponding target tracking state is generated for each target candidate.

[0007] Optionally, the generation of the multiple target state reversibility representations specifically includes: Based on the target tracking status, the target state vectors stored under multiple historical time indices of the corresponding target are read from the target historical state buffer according to the target identity identifier, and the target state vector sequence is formed in the order of time index. Based on the time index corresponding to each target state vector in the target state vector sequence, the observation data aligned with the time index is extracted from the observation sequence, and the platform state data aligned with the time index is extracted from the platform state sequence. For each target state vector in the target state vector sequence, the target state vector is combined with the observation data and platform state data extracted under the corresponding time index to generate a joint state representation corresponding to the target state vector; Perform combination processing on different target state vectors in the target state vector sequence to generate multiple joint state representations corresponding to each target state vector; Based on multiple joint state representations, the data is input into a deep learning model for training and optimization, generating multiple target state invertibility representations corresponding to different target state vectors.

[0008] Optionally, the generation of the target state reversibility constraint domain specifically includes: Based on multiple target state reversibility representations, the target state reversibility representations are aggregated according to the target identity identifier to form a set of target state reversibility representations corresponding to the same target identity identifier. The target state reversibility representations in the target state reversibility representation set are sorted according to their generation order to form a target state reversibility representation set with an ordered relationship. Accumulation processing is performed on the set of target state reversibility representations with sequential relationships, and each target state reversibility representation is incorporated into the accumulation result to form the accumulated target state reversibility representation result. Based on the accumulated target state reversibility characterization results, the value range of the target state reversibility characterization is characterized, and the allowable range of the target state reversibility characterization is determined. Based on the allowable range, construct the target state reversibility constraint domain.

[0009] Optionally, the generation of the reversibility verification result specifically includes: Based on multiple target state reversibility representations, select the target state reversibility representation corresponding to the current target state vector to generate the current target state reversibility representation; Map the current target state reversibility representation to the target state reversibility constraint domain to generate the positional relationship of the current target state reversibility representation in the target state reversibility constraint domain; Based on the positional relationship, a verification process is performed on the conformity between the current target state reversibility representation and the target state reversibility constraint domain, and a verification representation result is generated. Based on the verification representation results, a reversibility verification result corresponding to the reversibility representation of the current target state is generated.

[0010] Optionally, the generation of the target tracking task execution certificate specifically includes: Based on the reversibility verification result, the verification information corresponding to the current target identity is extracted, and the certificate verification input corresponding to the target identity is generated; Based on the certificate verification input, the executable status of the target tracking task is determined, and the target tracking task status determination result is generated; Map the target tracking task status determination result to the target tracking task execution certificate status, and generate a certificate status identifier; Based on the certificate status identifier, determine the target tracking task execution certificate status corresponding to the target identity identifier from the valid status, downgraded status, and revoked status; Bind the target identity identifier to the determined target tracking task execution certificate status to generate the target tracking task execution certificate corresponding to the target identity identifier.

[0011] Optionally, the state handling specifically includes: Obtain the target tracking task execution certificate corresponding to the target identity identifier, and parse the target tracking task execution certificate status; When the target tracking task execution certificate status is valid, the target status vector is updated based on the current target tracking status, generating an updated target status vector; Write the updated target state vector into the target history state buffer to generate the updated target history state buffer. When the target tracking task execution certificate status is downgraded, keep the target history state buffer unchanged and generate an unupdated target history state buffer. When the target tracking task execution certificate status is revoked, the target tracking task is terminated and a target tracking task termination result is generated. Based on the updated target history state buffer, the unupdated target history state buffer, or the target tracking task termination result, a target tracking task handling result corresponding to the target identity is generated.

[0012] The beneficial effects of this invention are: This invention proposes a deep learning-based dynamic target tracking method for unmanned aerial vehicles (UAVs), addressing the problems of insufficient target tracking accuracy and inflexible task execution control in existing technologies under complex environments. Most existing target tracking methods rely on traditional computer vision and filtering algorithms. While these methods can achieve basic target tracking under ideal conditions, their limitations become particularly pronounced when the UAV's attitude changes drastically, the environment is complex, or the observation perspective changes rapidly. Traditional algorithms cannot effectively cope with dynamic changes in the target's state, easily leading to mistracking, accumulation of state errors, and distortion of tracking results. Especially when the target is moving rapidly or in complex backgrounds, existing technologies fall short. Therefore, this invention introduces deep learning technology, based on the reversibility and dynamic changes of the target's state, to propose a more intelligent solution.

[0013] By using a deep learning model and combining observation data and platform state sequences, this invention can accurately capture the dynamic change characteristics of the target and generate a reversible representation of the target state. This representation can not only accurately reflect the change pattern of the target at different time steps, but also intelligently predict the target state, thereby avoiding the tracking errors caused by environmental changes in traditional methods. At the same time, this invention realizes intelligent management of the target tracking task through the reversibility constraint domain of the target state and the execution certificate mechanism. When the target tracking task execution certificate is in a valid state, the method can update the target state vector in real time and execute the task. In the degraded state, the historical state of the target is kept unchanged, ensuring the stability of the task. In the undo state, the tracking task is automatically terminated, thereby effectively avoiding the risk of mistracking.

[0014] The beneficial effects of this invention are that the introduction of deep learning greatly improves the accuracy and robustness of target tracking. Through an intelligent task execution control mechanism, the task execution strategy can be flexibly adjusted according to changes in the target state, ensuring that high tracking accuracy and stability can still be maintained in complex and dynamic environments. In addition, the verification mechanism based on the reversibility representation of the target state enhances the adaptability to changes in the target state and improves the operability and safety in real-time dynamic tasks. Finally, the use of a deep learning model for dynamic adjustment and target state update reduces the need for manual intervention and significantly improves task execution efficiency. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based UAV target dynamic tracking method proposed in this invention; Figure 2This is a schematic diagram illustrating the accumulation of target state invertibility representation and the construction of the target state invertibility constraint domain in a deep learning-based UAV target dynamic tracking method proposed in this invention. Figure 3 This is a schematic diagram illustrating the target state reversibility verification process of a deep learning-based UAV target dynamic tracking method proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A deep learning-based method for dynamic target tracking in unmanned aerial vehicles (UAVs) includes the following steps: During the flight of the UAV, a continuous image sequence is acquired and arranged according to the time index to generate an observation sequence. At the same time, the UAV attitude information and camera parameters are acquired and aligned according to the same time index to generate a platform status sequence. Based on the observation sequence and platform state sequence, a target candidate set is generated, the target is initialized, and a target tracking state is generated. The target tracking state includes a target identity identifier, a target state vector, and a target historical state buffer. Based on the target tracking state, the target state vector sequence is obtained from the target historical state buffer. The target state vector sequence, observation sequence and platform state sequence are input into the deep learning model to perform state representation inference and generate multiple target state invertibility representations corresponding to different target state vectors. Accumulate multiple target state reversibility representations to generate a target state reversibility constraint domain; Map the current target state reversibility representation to the target state reversibility constraint domain to generate reversibility verification results; Based on the reversibility verification results, a target tracking task execution certificate corresponding to the target identity is generated. The target tracking task execution certificate includes a valid status, a downgraded status, and a revoked status. The target tracking task is state-managed based on the target tracking task execution certificate. When it is in a valid state, the target state vector is updated and written to the target historical state buffer. When it is in a degraded state, the target historical state buffer is not updated. When it is in a revoked state, the target tracking task is terminated.

[0018] In this embodiment, the generation of the target tracking state specifically includes: Under each time index, target candidate generation processing is performed on the observation sequence, target regions in the observation sequence are detected and located, target regions that meet the detection conditions are extracted, and an initial target candidate set is generated. The initial target candidate set is associated with the platform state sequence. Based on the UAV attitude information and camera parameters in the platform state sequence, spatial consistency screening is performed on each target candidate in the initial target candidate set to remove target candidates that do not meet the spatial constraints and generate a filtered target candidate set. For each target candidate in the filtered target candidate set, target initialization processing is performed to assign a corresponding target identity identifier to each target candidate; Based on the spatial location of the target candidate in the observation sequence and the UAV attitude information and camera parameters in the platform state sequence, a corresponding target state vector is generated for each target identity. Using the target identity as an index, a corresponding target history state buffer is established for each target identity. The target state vectors generated under each time index are written into the corresponding target history state buffers in order of time index. The specific steps of writing to the corresponding target historical state buffer include: An independent target history state buffer is created for each target identity, storing all target state vectors corresponding to that target identity. Target state vectors generated at each time point are added sequentially to their corresponding target history state buffers according to their time index. During each write operation, the target state vectors are sorted by time index to ensure that the target state vectors stored in the buffer are arranged in chronological order. When the target history state buffer reaches a preset capacity, a caching strategy is executed, including deleting the oldest target state vector, using a sliding window approach, and retaining the target state vectors within the most recent time period. To ensure data integrity and consistency, a verification operation is performed when writing target state vectors to verify their validity, ensuring that the data stored in the target history state buffer is complete, correct, and corresponds to the target identity. If a conflict occurs when writing a new state vector to the target history state buffer, a fault tolerance mechanism is triggered, performing data repair and supplementation, and recording operation logs to ensure that the target history state buffer corresponding to the target identity is always in a valid state. The target identity identifier, target state vector, and target historical state buffer are bound together to generate target tracking state. When there are multiple target candidates in the target candidate set, a corresponding target tracking state is generated for each target candidate. The generation of the target tracking state specifically includes: Obtain the target identity identifier generated after target initialization, as the unique index identifier of the target tracking state; obtain the target state vector corresponding to the target identity identifier, as the current state representation in the target tracking state; obtain the target historical state buffer corresponding to the target identity identifier, as the historical state storage part in the target tracking state; combine the target identity identifier, target state vector, and target historical state buffer to generate the target tracking state corresponding to the target identity identifier; when there are multiple target candidates in the target candidate set, perform target identity identifier acquisition, target state vector acquisition, and target historical state buffer acquisition for each target candidate to generate multiple independent target tracking states.

[0019] In this embodiment, the generation of the reversibility representations of the multiple target states specifically includes: Based on the target tracking status, the target state vectors stored under multiple historical time indices of the corresponding target are read from the target historical state buffer according to the target identity identifier, and the target state vector sequence is formed in the order of time index. Based on the time index corresponding to each target state vector in the target state vector sequence, the observation data aligned with the time index is extracted from the observation sequence, and the platform state data aligned with the time index is extracted from the platform state sequence. For each target state vector in the target state vector sequence, the target state vector is combined with the observation data and platform state data extracted under the corresponding time index to generate a joint state representation corresponding to the target state vector; Perform combination processing on different target state vectors in the target state vector sequence to generate multiple joint state representations corresponding to each target state vector; Based on multiple joint state representations, the data is input into a deep learning model for training and optimization, generating multiple target state invertibility representations corresponding to different target state vectors. The generation of the reversibility representations of the multiple target states specifically includes: Multiple joint state representations are distinguished according to the target identity, forming a set of joint state representations corresponding to the same target identity. The joint state representations in the set are arranged according to the order of their corresponding target state vectors, forming a sequence of joint state representations with an order relationship. The sequence of joint state representations is input into a deep learning model, and state representation inference processing is performed on the sequence to generate intermediate state representation results corresponding to each joint state representation in the sequence. Based on the intermediate state representation results, state representation information reflecting the stability of target state evolution and state backtracking characteristics is extracted to generate target state invertibility representations corresponding to each target state vector. The generated target state invertibility representations are collected to form multiple target state invertibility representations corresponding to different target state vectors.

[0020] In this embodiment, the generation of the target state reversibility constraint domain specifically includes: Based on multiple target state reversibility representations, the target state reversibility representations are aggregated according to the target identity identifier to form a set of target state reversibility representations corresponding to the same target identity identifier. The target state reversibility representations in the target state reversibility representation set are sorted according to their generation order to form a target state reversibility representation set with an ordered relationship. Accumulation processing is performed on the set of target state reversibility representations with sequential relationships, and each target state reversibility representation is incorporated into the accumulation result to form the accumulated target state reversibility representation result. Based on the accumulated target state reversibility characterization results, the value range of the target state reversibility characterization is characterized, and the allowable range of the target state reversibility characterization is determined. The generation of the allowed range specifically includes: Obtain multiple target state reversibility representations, calculate their minimum and maximum values, and generate the upper and lower boundaries of the target state reversibility representations; based on the upper and lower boundaries, expand the fluctuation range to ensure that the target state reversibility representations fluctuate within the normal range; combine the fluctuation range and the upper and lower boundaries to determine the actual allowable range of the target state reversibility representations. Based on the allowable range, construct the target state reversibility constraint domain.

[0021] In this embodiment, the generation of the reversibility verification result specifically includes: Based on multiple target state reversibility representations, select the target state reversibility representation corresponding to the current target state vector to generate the current target state reversibility representation; Map the current target state reversibility representation to the target state reversibility constraint domain to generate the positional relationship of the current target state reversibility representation in the target state reversibility constraint domain; The generation of the positional relationship specifically includes: When mapping the current target state reversibility representation to the target state reversibility constraint domain, the current target state reversibility representation and the corresponding target state reversibility constraint domain are obtained; the current target state reversibility representation is compared with the constraint boundary of the target state reversibility constraint domain to determine the relative position of the current target state reversibility representation relative to the target state reversibility constraint domain; the relative position is used to characterize the relationship state between the current target state reversibility representation and the target state reversibility constraint domain, including being located inside the target state reversibility constraint domain, located at the boundary of the target state reversibility constraint domain, or located outside the target state reversibility constraint domain; based on the relationship state, the positional relationship of the current target state reversibility representation in the target state reversibility constraint domain is generated; Based on the positional relationship, a verification process is performed on the conformity between the current target state reversibility representation and the target state reversibility constraint domain, and a verification representation result is generated. The generation of the verification and characterization results specifically includes: Obtain the positional relationship of the current target state reversibility representation within the target state reversibility constraint domain; based on the positional relationship, determine whether the current target state reversibility representation and the target state reversibility constraint domain satisfy a preset constraint consistency condition; when the positional relationship indicates that the current target state reversibility representation is located inside the target state reversibility constraint domain, generate the corresponding constraint consistency representation; when the positional relationship indicates that the current target state reversibility representation is located at the boundary of the target state reversibility constraint domain, generate the corresponding boundary consistency representation; when the positional relationship indicates that the current target state reversibility representation is located outside the target state reversibility constraint domain, generate the corresponding constraint inconsistency representation; use the generated constraint consistency representation, boundary consistency representation, or constraint inconsistency representation as the verification representation result; Based on the verification representation results, generate reversibility verification results corresponding to the reversibility representation of the current target state; The generation of the reversibility verification result specifically includes: Obtain the verification representation results, which include constraint consistency representation, boundary consistency representation, or constraint inconsistency representation. Based on the constraint consistency representation, boundary consistency representation, and constraint inconsistency representation, determine whether the target state reversibility representation is within the valid range of the target state reversibility constraint domain. When the target state reversibility representation is within the target state reversibility constraint domain, generate a reversibility verification result of "valid state". When the target state reversibility representation is at the boundary of the target state reversibility constraint domain, generate a reversibility verification result of "degraded state". When the target state reversibility representation is outside the target state reversibility constraint domain, generate a reversibility verification result of "cancelled state".

[0022] In this embodiment, the generation of the target tracking task execution certificate specifically includes: Based on the reversibility verification result, the verification information corresponding to the current target identity is extracted, and the certificate verification input corresponding to the target identity is generated; Based on the certificate verification input, the executable status of the target tracking task is determined, and the target tracking task status determination result is generated; Map the target tracking task status determination result to the target tracking task execution certificate status, and generate a certificate status identifier; The generation of the certificate status identifier specifically includes: When the target tracking task status determination result indicates that the target tracking task meets the conditions for continued execution, the certificate status is marked as "valid"; when the target tracking task status determination result indicates that the target tracking task is at the executable boundary condition, the certificate status is marked as "degraded"; when the target tracking task status determination result indicates that the target tracking task does not meet the conditions for continued execution, the certificate status is marked as "revoked". Based on the certificate status identifier, determine the target tracking task execution certificate status corresponding to the target identity identifier from the valid status, downgraded status, and revoked status; The determination of the target tracking task execution certificate status specifically includes: Obtain the certificate status identifier corresponding to the target identity identifier; when the certificate status identifier is "valid", determine the target tracking task execution certificate status corresponding to the target identity identifier as "valid"; when the certificate status identifier is "downgraded", determine the target tracking task execution certificate status corresponding to the target identity identifier as "downgraded"; when the certificate status identifier is "revoked", determine the target tracking task execution certificate status corresponding to the target identity identifier as "revoked". Bind the target identity identifier to the determined target tracking task execution certificate status to generate the target tracking task execution certificate corresponding to the target identity identifier.

[0023] In this embodiment, the state handling specifically includes: Obtain the target tracking task execution certificate corresponding to the target identity identifier, and parse the target tracking task execution certificate status; When the target tracking task execution certificate status is valid, the target status vector is updated based on the current target tracking status, generating an updated target status vector; The generation of the updated target state vector specifically includes: The system acquires the current target tracking status, including the target identifier, the current target state vector, and the target historical state buffer. Based on the target identifier and the current target state vector, it extracts the corresponding target state vector sequence from the target historical state buffer. Based on the target state vector sequence and the observation data at the current time index, it extracts the spatial and temporal features of the target using a convolutional neural network and updates the current target state vector. It uses a deep learning model for smoothing to remove noise and improve the accuracy of the target state vector. Based on the updated target state vector and the platform state sequence, it generates a new target state vector. Write the updated target state vector into the target history state buffer to generate the updated target history state buffer. When the target tracking task execution certificate status is downgraded, keep the target history state buffer unchanged and generate an unupdated target history state buffer. When the target tracking task execution certificate status is revoked, the target tracking task is terminated and a target tracking task termination result is generated. Based on the updated target history state buffer, the unupdated target history state buffer, or the target tracking task termination result, a target tracking task handling result corresponding to the target identity is generated.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a UAV target dynamic tracking task. This task mainly involves using UAVs for aerial patrol, continuously acquiring and processing images to accurately track a designated target, and making real-time adjustments and controls based on the target's dynamic characteristics when the target's state changes. A typical UAV target tracking scenario was selected, with the background being a complex urban environment and the target being a vehicle moving on the ground. Under different flight altitudes and speeds, the accuracy of target tracking and the stability of task execution became the key challenges.

[0025] In this scenario, the drone pilot controls the drone through a remote control device, continuously collecting real-time data during the flight. This data includes image sequences from the drone, attitude information of the flight platform, and intrinsic and extrinsic parameter data from the camera. Through this data, the dynamic behavior of the target and environmental information can be acquired in real time. The image sequences are mainly used to identify ground targets, while the platform status sequences reflect the changes in the aircraft's own position and attitude, helping to better judge the dynamic changes of the target in complex environments. Throughout the flight, the core task of target tracking is to ensure that the drone can continuously and accurately identify and track the target, especially when encountering obstacles or significant changes in flight attitude during the flight, the tracking task can still maintain stability.

[0026] In this scenario, the UAV target dynamic tracking method of the present invention is applied through the following steps: First, image sequences are continuously acquired, and based on changes in the target area, a deep learning model is used to identify the target and determine its trajectory. By combining the platform's state information, including attitude data and camera parameters, the target's position in space is determined. Furthermore, through the inference capability of the deep learning model, the evolution law of the target state is analyzed to generate a reversible representation of the target state. When the target undergoes rapid changes or the complexity of the environment increases, the method can identify the trend of the target state change and determine whether the target state has the conditions to continue execution, thus avoiding mistracking and the accumulation of erroneous states.

[0027] This method, through the introduction of deep learning, enables the target tracking process to not only rely on traditional geometric and motion models, but also to intelligently adapt to the dynamic changes of the target through model learning, optimize the prediction of the target state, and react in a timely manner when the target state undergoes significant changes. It generates a new target state reversibility representation and maps it to the target state reversibility constraint domain for verification. This process ensures that when the target state changes significantly, it can determine whether to continue tracking the target based on the reversibility verification result. If the target state is found to be inconsistent with the tracking conditions, the task is automatically terminated or the tracking strategy is adjusted.

[0028] Furthermore, this invention provides a scheme for controlling the task execution state through a target tracking task execution certificate. During task execution, a target tracking task execution certificate is generated based on the generated reversibility verification result. This certificate includes valid, downgraded, and revoked states. If the target state still meets the tracking requirements, the execution certificate remains valid and continues to update the target state. If the target state changes but is still within an acceptable range, the execution certificate enters a downgraded state, keeping the target state buffer unchanged. If the target state does not meet the tracking requirements, the target tracking task is revoked, and task execution is automatically terminated. This intelligent task control mechanism ensures that the target tracking task can be executed under appropriate conditions in any environment, avoiding misoperations and distorted results in traditional methods.

[0029] By applying this method, UAVs can track targets in real time in complex and dynamic environments, adapting to various challenges such as changes in flight altitude, target position, and perspective. At the same time, intelligent control of mission execution can greatly reduce errors and computational costs, improving the accuracy and robustness of target tracking. During testing, the method was verified multiple times under different flight scenarios. Although the target position, speed, and direction changed, this invention, through the combination of deep learning models and mission execution certificates, was always able to maintain efficient target tracking capabilities. Especially in situations with significant environmental interference or unstable target dynamics, it can stably execute missions, avoiding the error accumulation and distortion problems of traditional tracking methods.

[0030] Table 1. Performance Comparison of Different Target Dynamic Tracking Methods

[0031] As shown in Table 1, traditional target tracking methods in actual UAV flight scenarios mainly rely on current observation information or limited short-term historical information for state updates. When the target is occluded, scale changes, perspective changes abruptly, or the UAV attitude is rapidly adjusted, the tracking state is prone to cumulative deviations, resulting in an average tracking accuracy of around 0.86. At the same time, the target loss rate and mistracking rate are relatively high. These methods have certain effectiveness on a single frame or short time scale, but in continuous flight missions, due to the lack of characterization and constraint on the reversibility of target state evolution, once the erroneous state is written into the historical process, it tends to continue to amplify, thus affecting the stability of subsequent tracking. In contrast, the method of this invention introduces a target historical state buffer and combines a deep learning model to jointly infer the state representation of the target state vector sequence, observation sequence, and platform state sequence. This makes the target state update no longer dependent on information at a single moment, but based on continuous time state evolution, thereby improving the average tracking accuracy to 0.90. Moreover, this improvement process exhibits stable and gradual characteristics, which is consistent with the performance characteristics in real flight missions.

[0032] Regarding target loss rate and false tracking rate, traditional methods tend to misjudge short-term anomalies as the true state of the target when the target moves rapidly or when background interference is enhanced, resulting in a target loss rate of 0.14 and a false tracking rate of 0.11. However, the method of this invention constructs a target state reversibility constraint domain to verify the reversibility representation of the current target state, preventing abnormal states that do not conform to historical evolution patterns from directly entering the state update process. This effectively suppresses the propagation of abnormal states. This mechanism reduces the target loss rate to 0.09 and the false tracking rate to 0.07. Although the numerical variation is controlled within a reasonable range, it can significantly reduce the cumulative probability of tracking failure in continuous tasks, demonstrating higher engineering practical value.

[0033] From the perspective of tracking state stability, traditional methods have limited ability to maintain stable tracking over continuous time, with a tracking state stability retention rate of 0.82 and a continuous effective tracking duration ratio of 0.79. This indicates that there are still many state fluctuations within a relatively long time window. The method of this invention accumulates the reversibility representations of multiple target states to form a target state reversibility constraint domain, ensuring that the target state update process is always limited to a reasonable evolution range. Combined with the target tracking task execution certificate mechanism, when the state approaches the unstable boundary, it promptly triggers downgrade or termination, preventing unreliable states from being continuously written into the historical buffer. This increases the tracking state stability retention rate to 0.88 and the continuous effective tracking duration ratio to 0.85. This improvement does not rely on a single judgment result, but rather stems from long-term constraints and feedback control on the target state evolution process.

[0034] Furthermore, the difference between the two methods under long-term operating conditions can be further observed from the frequency of abnormal tracking state triggers. Traditional methods, lacking a clear state verification and handling mechanism, often only passively correct abnormal states in subsequent stages, resulting in an abnormal trigger frequency of around 0.18. In contrast, the method of this invention introduces reversibility verification results before state updates and generates a target tracking task execution certificate based on the verification results, thereby implementing hierarchical control over state update behavior. This allows abnormal states to be identified and isolated early, reducing the abnormal trigger frequency to 0.12. This result demonstrates that this invention, while maintaining algorithm stability, can effectively reduce the accumulation of risks during long-term operation and possesses application characteristics more suitable for continuous dynamic target tracking in complex flight environments.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for dynamic target tracking in unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: During the flight of the UAV, a continuous image sequence is acquired and arranged according to the time index to generate an observation sequence. At the same time, the UAV attitude information and camera parameters are acquired and aligned according to the same time index to generate a platform status sequence. Based on the observation sequence and platform state sequence, a target candidate set is generated, the target is initialized, and a target tracking state is generated. The target tracking state includes a target identity identifier, a target state vector, and a target historical state buffer. Based on the target tracking state, the target state vector sequence is obtained from the target historical state buffer. The target state vector sequence, observation sequence and platform state sequence are input into the deep learning model to perform state representation inference and generate multiple target state invertibility representations corresponding to different target state vectors. Accumulate multiple target state reversibility representations to generate a target state reversibility constraint domain; Map the current target state reversibility representation to the target state reversibility constraint domain to generate reversibility verification results; Based on the reversibility verification results, a target tracking task execution certificate corresponding to the target identity is generated. The target tracking task execution certificate includes a valid status, a downgraded status, and a revoked status. The target tracking task is state-managed based on the target tracking task execution certificate. When it is in a valid state, the target state vector is updated and written to the target historical state buffer. When it is in a degraded state, the target historical state buffer is not updated. When it is in a revoked state, the target tracking task is terminated.

2. The method for dynamic target tracking of a UAV based on deep learning according to claim 1, characterized in that, The generation of the target tracking state specifically includes: Under each time index, target candidate generation processing is performed on the observation sequence, target regions in the observation sequence are detected and located, target regions that meet the detection conditions are extracted, and an initial target candidate set is generated. The initial target candidate set is associated with the platform state sequence. Based on the UAV attitude information and camera parameters in the platform state sequence, spatial consistency screening is performed on each target candidate in the initial target candidate set to eliminate target candidates that do not meet the spatial constraints and generate a filtered target candidate set. For each target candidate in the filtered target candidate set, target initialization processing is performed to assign a corresponding target identity identifier to each target candidate; Based on the spatial location of the target candidate in the observation sequence and the UAV attitude information and camera parameters in the platform state sequence, a corresponding target state vector is generated for each target identity. Using the target identity as an index, a corresponding target history state buffer is established for each target identity. The target state vectors generated under each time index are written into the corresponding target history state buffers in the order of the time index. The target identity identifier, target state vector, and target historical state buffer are bound together to generate target tracking state. When there are multiple target candidates in the target candidate set, a corresponding target tracking state is generated for each target candidate.

3. The method for dynamic target tracking of a UAV based on deep learning according to claim 1, characterized in that, The generation of the reversibility representations of the multiple target states specifically includes: Based on the target tracking status, the target state vectors stored under multiple historical time indices of the corresponding target are read from the target historical state buffer according to the target identity identifier, and the target state vector sequence is formed in the order of time index. Based on the time index corresponding to each target state vector in the target state vector sequence, the observation data aligned with the time index is extracted from the observation sequence, and the platform state data aligned with the time index is extracted from the platform state sequence. For each target state vector in the target state vector sequence, the target state vector is combined with the observation data and platform state data extracted under the corresponding time index to generate a joint state representation corresponding to the target state vector; Perform combination processing on different target state vectors in the target state vector sequence to generate multiple joint state representations corresponding to each target state vector; Based on multiple joint state representations, the data is input into a deep learning model for training and optimization, generating multiple target state invertibility representations corresponding to different target state vectors.

4. The method for dynamic target tracking of a UAV based on deep learning according to claim 1, characterized in that, The generation of the target state reversibility constraint domain specifically includes: Based on multiple target state reversibility representations, the target state reversibility representations are aggregated according to the target identity identifier to form a set of target state reversibility representations corresponding to the same target identity identifier. The target state reversibility representations in the target state reversibility representation set are sorted according to their generation order to form a target state reversibility representation set with an ordered relationship. Accumulation processing is performed on the set of target state reversibility representations with sequential relationships, and each target state reversibility representation is incorporated into the accumulation result to form the accumulated target state reversibility representation result. Based on the accumulated target state reversibility characterization results, the value range of the target state reversibility characterization is characterized, and the allowable range of the target state reversibility characterization is determined. Based on the allowable range, construct the target state reversibility constraint domain.

5. The method for dynamic tracking of unmanned aerial vehicle targets based on deep learning according to claim 1, characterized in that, The generation of the reversibility verification result specifically includes: Based on multiple target state reversibility representations, select the target state reversibility representation corresponding to the current target state vector to generate the current target state reversibility representation; Map the current target state reversibility representation to the target state reversibility constraint domain to generate the positional relationship of the current target state reversibility representation in the target state reversibility constraint domain; Based on the positional relationship, a verification process is performed on the conformity between the current target state reversibility representation and the target state reversibility constraint domain, and a verification representation result is generated. Based on the verification representation results, a reversibility verification result corresponding to the reversibility representation of the current target state is generated.

6. The method for dynamic target tracking of a UAV based on deep learning according to claim 1, characterized in that, The generation of the target tracking task execution certificate specifically includes: Based on the reversibility verification result, the verification information corresponding to the current target identity is extracted, and the certificate verification input corresponding to the target identity is generated; Based on the certificate verification input, the executable status of the target tracking task is determined, and the target tracking task status determination result is generated; Map the target tracking task status determination result to the target tracking task execution certificate status, and generate a certificate status identifier; Based on the certificate status identifier, determine the target tracking task execution certificate status corresponding to the target identity identifier from the valid status, downgraded status, and revoked status; Bind the target identity identifier to the determined target tracking task execution certificate status to generate the target tracking task execution certificate corresponding to the target identity identifier.

7. The method for dynamic target tracking of a UAV based on deep learning according to claim 1, characterized in that, The status handling specifically includes: Obtain the target tracking task execution certificate corresponding to the target identity identifier, and parse the target tracking task execution certificate status; When the target tracking task execution certificate status is valid, the target status vector is updated based on the current target tracking status, generating an updated target status vector; Write the updated target state vector into the target history state buffer to generate the updated target history state buffer. When the target tracking task execution certificate status is downgraded, keep the target history state buffer unchanged and generate an unupdated target history state buffer. When the target tracking task execution certificate status is revoked, the target tracking task is terminated and a target tracking task termination result is generated. Based on the updated target history state buffer, the unupdated target history state buffer, or the target tracking task termination result, a target tracking task handling result corresponding to the target identity is generated.

Images