Multi-rotor aircraft cooperative target tracking method based on adaptive attitude correction

Abstract

The invention provides a multi-rotor aircraft cooperative target tracking method based on adaptive attitude correction, and the method comprises the steps: carrying out the modeling of a multi-rotor aircraft cooperative target tracking task, carrying out the modeling of a precise target tracking task and a global visual angle coverage task, and determining the cooperative relation between aircrafts and a task target; designing a self-adaptive attitude correction mechanism, and re-capturing the target based on attitude information of other aircrafts in a precise target tracking task; designing a collaborative mechanism based on view angle and distance optimization, optimizing relative positions and view angle distribution between the aircrafts through weighted combination of local rewards and global rewards, and completing a global view angle coverage task; multi-rotor aircraft cooperative target tracking based on adaptive attitude correction is realized by adopting multiple aircrafts, an adaptive attitude correction mechanism and a cooperative mechanism based on view angle and distance optimization at the same time. According to the method, the tracking target loss condition can be effectively handled in a complex environment, and the good continuous tracking capability is kept.

Description

Technical Field

[0001] The present invention relates to a collaborative target tracking method, in particular to a multi-rotor aircraft collaborative target tracking method based on adaptive posture correction. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] In recent years, with the rapid development of UAV (Unmanned Aerial Vehicle) technology, the application of UAVs in the fields of industry, environmental monitoring, agriculture, emergency rescue, etc. has become increasingly widespread, and has gradually become an important tool for improving mission efficiency and execution accuracy. For example, in the field of environmental monitoring, UAVs are widely used in forest fire prevention [1], air quality monitoring and water quality sampling; in agriculture, UAVs can be used for precise pesticide application and crop health monitoring; in emergency rescue, UAVs can quickly reach disaster areas and perform search and rescue missions, improving response speed and safety. However, with the increasing complexity of mission scenarios, the coverage and capabilities of a single UAV in performing missions are often limited, making it difficult to meet the needs of large-scale or high-difficulty real-time monitoring. Against this background, multi-UAV collaborative technology has emerged, which uses the collaborative cooperation of multiple UAVs to improve the flexibility and efficiency of mission execution. Multi-UAV systems can not only achieve wider-range monitoring, but also respond and adapt to mission changes more quickly in complex dynamic environments.

[0004] Existing active target tracking technologies are mostly applied to fixed-camera scenarios, continuously tracking a target by adjusting the camera and its angle. However, cameras have a less flexible motion model than drones. Once a target leaves the camera's coverage area, tracking is lost. Even with multiple cameras working together, the area they can maintain tracking is relatively limited, making them unsuitable for complex tracking scenarios and tasks.

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a multi-rotor aircraft collaborative target tracking method based on adaptive posture correction in response to the shortcomings of the existing technology.

[0007] In order to solve the above technical problems, the present invention discloses a multi-rotor aircraft collaborative target tracking method based on adaptive posture correction, comprising the following steps:

[0008] Step 1: Model the multi-rotor aircraft collaborative target tracking task, modeling it as a precise target tracking task and a global view coverage task, and determine the collaborative relationship and task objectives between the aircraft;

[0009] Step 2: Design an adaptive attitude correction mechanism to recapture the target based on the attitude information of other aircraft in the precise target tracking task;

[0010] Step 3: Design a collaborative mechanism based on perspective and distance optimization. Through a weighted combination of local rewards and global rewards, optimize the relative positions and perspective distribution between aircraft to achieve the global perspective coverage task.

[0011] Step 4: Use multiple aircraft to simultaneously adopt the adaptive attitude correction mechanism and the collaborative mechanism based on view angle and distance optimization to achieve multi-rotor aircraft collaborative target tracking based on adaptive attitude correction.

[0012] Furthermore, the multi-rotor aircraft cooperative target tracking task described in step 1 is modeled, including:

[0013] Step 1-1: The task of multi-aircraft collaborative active tracking of a single target is divided into the task of precise target tracking and the task of global view coverage, as follows:

[0014] Precision tracking missions, where when a single aircraft loses its target, other aircraft assist in calibration;

[0015] Global perception tasks, where a single aircraft considers the distance and view distribution between it and other aircraft when choosing its own actions;

[0016] In steps 1 and 2, the precise target tracking task and the global view coverage task are modeled respectively.

[0017] Furthermore, the precise target tracking task and the global view coverage task described in steps 1-2 are modeled separately as follows:

[0018] Step 1-2-1: Model the precise target tracking task as a multi-agent deep reinforcement learning problem as follows:

[0019] Assume the input state of the i-th aircraft is s i , contains the position information and attitude angle information of the aircraft, indicating the globally perceived environmental state;

[0020] Based on the deep reinforcement learning method, the training strategy network π(a|s i ), used to generate action strategies, that is, at each time step t, the strategy network outputs the action a of the i-th aircraft i , used to adjust the attitude and position of the aircraft so that the camera is re-aimed at the target direction;

[0021] Build a global perception network, rely on the location information provided by other aircraft to infer the possible location of the target, and adjust the aircraft's own posture and position accordingly, so that the camera is aimed at the potential target area;

[0022] Step 1-2-2 covers the global perspective task as follows:

[0023] Under the coordinated tracking of multiple aircraft, the mission goal is to achieve all-round coverage of the target.

[0024] Furthermore, the design of the adaptive posture correction mechanism described in step 2 includes:

[0025] Vision strategy module, controller module and posture strategy module;

[0026] At each moment, the observation information and attitude information of each aircraft are input into the visual strategy module, controller module and attitude strategy module respectively;

[0027] The self-vision strategy module and the self-attitude strategy module generate candidate tracking instructions respectively, which are executed by the controller module. If the aircraft does not lose the target, the candidate tracking instructions generated by the vision strategy module are executed; otherwise, the candidate tracking instructions generated by the attitude strategy module are executed.

[0028] Furthermore, the controller module includes:

[0029] The observed image features processed by the convolutional neural network (CNN) and the long short-term memory (LSTM) network are received and input into the first network consisting of two fully connected layers for state classification, which is represented as follows:

[0030] h1=ReLU(W1·x+b1)

[0031] h2=ReLU(W2·h1+b2)

[0032] out=(W3·h2+b3)

[0033] Among them, x represents the input observation image feature vector, w1 and w2 are the weight matrices of the fully connected layer, b1 and b2 are bias terms, ReLU represents the activation function, W3 and b3 represent the weight matrix and bias term of the final output layer, and the final output of the first network is out, which is a vector of shape [N, 2], denoted as:

[0034] out[i]=[z1,z2]

[0035] Among them, out[i] is the status output of the i-th aircraft, z1 corresponds to the original score of aircraft i successfully tracking the target, and z2 corresponds to the original score of aircraft losing the target;

[0036] After the normalized exponential function Softmax operation, the probability P is obtained p , which is expressed as follows:

[0037]

[0038] Where P1 represents the probability of successfully tracking the target, P2 represents the probability of losing the target, P1+P2=1;

[0039] Through the maximum value operation, it is expressed as follows:

[0040] index[i]=argmaxP p

[0041] Get the classification result index[i] of the current state, that is, complete the judgment of whether the i-th aircraft has lost the target.

[0042] Furthermore, the posture strategy module includes:

[0043] Assume the position information of the i-th aircraft and the camera angle p i Expressed as:

[0044]

[0045] Among them, (x i ,y i ,z i ) is the position coordinate of the aircraft, and θ i Respectively represent the yaw angle and pitch angle of the camera on the aircraft;

[0046] Normalize the position information of the aircraft and the camera angle to obtain the normalized pose information combined feature vector c of the i-th aircraft i , which is expressed as follows:

[0047] c i =(p″ i ,sin yaw ,cos yaw ,sin pitch ,cos pitch )

[0048] Among them, p″ i is the normalized position information of the aircraft, sin yaw ,cos yaw ,sin pitch ,cos pitch is the normalized angle information, which are expressed as follows:

[0049]

[0050] Among them, (d x ,d y ,d z ) is the mean of the initial positions of all aircraft, σ is the scale factor, i.e. the maximum distance of the position offset of all aircraft;

[0051] Combined with the controller's output index[i], the input feature gf of the posture strategy module is obtained i , which is expressed as follows:

[0052] gf i =(c i ,index[i])

[0053] The input feature gf of the i-th aircraft i The input is fed into the GRU network for temporal information processing, which is expressed as follows:

[0054] output t =GRU(gf i ,h t-1 )

[0055] Among them, h t-1 is the hidden state at time step t-1, output t is the output vector at time step t;

[0056] Output vector output t The input is sent to the policy network, which uses the Actor-Critic structure to learn and optimize the drone control strategy and generate correction instructions as candidate tracking instructions output by the posture policy module.

[0057] Furthermore, the visual strategy module includes:

[0058] The observation image of the aircraft is subjected to feature extraction by the convolutional neural network (CNN) to obtain the target motion features;

[0059] Use the long short-term memory network (LSTM) to capture the dynamic changes of the target in the time dimension and obtain the time change features;

[0060] The target motion features and time change features are combined and input into the visual strategy module, which outputs the candidate vision-based tracking instructions for the aircraft.

[0061] Furthermore, the design described in step 3 is based on a synergistic mechanism of view angle and distance optimization, including:

[0062] Assume that the total reward obtained by the i-th aircraft at time t is R i (t), is expressed as follows:

[0063] R i (t) = α·R local,i (t)+β·R global (t)

[0064] Among them, R local,i (t) represents the independent tracking reward of the i-th aircraft, i.e., the local reward, R global (t) is the global reward, α and β are the weight factors of local and global rewards respectively;

[0065] The total reward is used to adjust network parameters for deep reinforcement learning methods in precise target tracking tasks.

[0066] Furthermore, the local reward R described in step 3 local,i (t), is calculated as follows:

[0067]

[0068] Among them, α(t), β(t) and d(t) represent the horizontal offset, vertical offset and target distance between the target and the center of the aircraft’s field of view, respectively. max , β max and d max is the corresponding normalization coefficient.

[0069] Furthermore, the global reward R described in step 3 global (t) is a linear combination of the viewing angle reward and the distance reward, expressed as follows:

[0070] R global (t) = R global，angle (t)+R global，distance (t)

[0071] Among them, the perspective reward R global，angle (t), is calculated as follows:

[0072]

[0073] Among them, γ angle is the weight factor of the view distribution reward, α angle is the angle error normalization coefficient, L angle is the view distribution optimization loss function, expressed as:

[0074]

[0075] Where the desired angle is θ target , the angle between the perspectives of the i-th and j-th aircraft is θ ij ;

[0076] Distance Reward R global，distance (t), is calculated as follows:

[0077]

[0078] Among them, γ distance is the weight factor of the distance reward, α distance is the distance error normalization coefficient, the distance loss function L distance , expressed as:

[0079]

[0080] The expected distance is d target , the distance between the i-th and j-th aircraft is d ii (t).

[0081] Beneficial effects:

[0082] 1. The adaptive attitude correction and view distance optimization method proposed in this invention improves the target tracking capability of the multi-aircraft system in complex environments.

[0083] 2. The present invention ensures all-round coverage by optimizing the relative positions and viewing angle distribution between aircraft, and effectively addresses situations where an aircraft loses its target due to occlusion or other reasons by sharing position information between aircraft.

[0084] 3. The present invention helps an aircraft that has lost its target to recapture the target by sharing the position information of the aircraft, achieves continuous tracking, and improves the robustness of a single aircraft in a complex environment.

[0085] 4. The present invention also introduces an optimization strategy based on perspective and distance, comprehensively considering the relative positions of the aircraft and the field of view coverage. By rationally designing the control strategy and action selection, it avoids the problem of perspective overlap caused by the close distance between aircraft and the tracking blind spots caused by uneven distribution, ensuring that multiple aircraft can collaboratively complete the tracking task with the optimal perspective and distance.

[0086] 5. The method proposed in the present invention enables multi-rotor aircraft to effectively deal with the loss of tracking targets through a collaborative tracking mechanism in complex environments and maintain good continuous tracking capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0088] Figure 1 It is a flow chart of the adaptive posture correction mechanism of the present invention.

[0089] Figure 2Schematic diagram of the virtual 3D environment used in the present invention.

[0090] Figure 3 It is a schematic diagram of a virtual 3D environment for simulating a display world used in the present invention.

[0091] Figure 4 It is a schematic diagram of the model of the adaptive posture correction method proposed in the present invention.

[0092] Figure 5 Schematic diagram of the viewing angle distance optimization method proposed in the present invention.

[0093] Figure 6 It is a schematic diagram of the simulation test tracking reward results conducted by the present invention.

[0094] Figure 7 It is a schematic diagram of the tracking success rate results of the simulation test performed by the present invention.

[0095] Figure 8 It is a 3D scatter diagram of the tracking performance indicators of the simulation test performed by the present invention.

[0096] Figure 9 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0097] The overall idea of the present invention is as follows: combined with active target tracking technology, the multi-UAV system can give full play to the flexibility and adaptability of UAVs. The core concept of active target tracking is that multiple UAVs equipped with cameras can autonomously plan paths and adjust flight attitudes and camera angles according to the real-time motion trajectory of the target and changes in the surrounding environment to ensure that the target is always under effective monitoring. This method not only reduces the risk of mistracking due to visual angle limitations in traditional passive tracking, but also can make efficient adjustments in complex environments to ensure continuous tracking of the target. By autonomously planning paths and coordinating actions, the multi-UAV system can flexibly respond to complex and uncertain dynamic environments, especially when the target changes rapidly or the environmental interference is large, and exhibits better performance than a single UAV. Therefore, active target tracking technology based on multi-UAV collaboration has become a very promising research direction.

[0098] This invention proposes a multi-UAV collaborative tracking solution based on adaptive attitude correction and perspective and distance optimization by designing scientific and reasonable collaborative tasks. Adaptive attitude correction helps UAVs that are blocked or lost due to target recapture targets by sharing UAV attitude information; perspective and distance optimization ensures that each UAV completes the tracking task with the optimal perspective and distance through reasonable control strategies and action selection. Through the above design, the overall tracking performance of the system is improved, and more complex active tracking tasks can be completed.

[0099] The technical solutions proposed by the present invention are as follows:

[0100] A method for collaborative active target tracking of multirotor aircraft based on adaptive attitude correction and view distance optimization includes the following three steps:

[0101] Step (1) For the collaborative active target tracking of multi-rotor aircraft, in order to cope with the tracking challenges in complex environments, the problems that may arise in the collaborative process are analyzed, and how to solve these obstacles and challenges through reasonable mechanism design is studied, so as to model and realize the collaborative active target tracking of multi-rotor aircraft.

[0102] In step (1), the goal of the present invention is to model and implement the problem of collaborative active target tracking of multi-rotor aircraft. More specifically, multiple aircraft collaborate to actively track a single target. Taking into account the need for continuous tracking, the problem is modeled as a multi-agent collaborative optimization problem, which can be specifically divided into two tasks: precise target tracking and global perspective coverage.

[0103] For the precise target tracking task, it mainly deals with the situation in which the aircraft may lose the target due to occlusion in the environment or sudden acceleration of the target. Therefore, the present invention models it as a multi-agent deep reinforcement learning problem to enhance the system's tracking ability through collaboration. In this model, the input state s of aircraft i is i Contains the position information and attitude angle information of all aircraft to represent the global perceived environment state. Based on the deep reinforcement learning method, the policy network π(a|s i ) is trained to generate action policies. At each time step t, the policy network outputs an action a i , which is used to adjust the attitude and position of the aircraft so that the camera can be re-aimed at the target direction to ensure that the target remains within the monitoring range. This modeling fully considers the actual challenges in the active tracking mission of aircraft, that is, the target may temporarily leave the field of view of a certain aircraft due to obstacles or sudden speed changes. To address this problem, a global perception network is built by sharing real-time position information and attitude angles between aircraft. When an aircraft loses its target, it can rely on the position information provided by other aircraft to infer the possible location of the target, and adjust its own attitude and position accordingly to aim the camera at the potential target area. This multi-agent collaborative perception method successfully expands the local perception capability of a single aircraft to the global perception capability of a multi-aircraft system. Through this modeling strategy, the impact of target loss can be mitigated while enhancing the robustness of the entire system in complex environments.

[0104] For global perception tasks, a significant advantage of multi-vehicle systems over single-vehicle systems is that the increased number of aircraft provides more perspective options and a wider field of view within the tracking environment. This feature enables multi-vehicle systems to perform better in complex environments while improving the stability and efficiency of mission completion. Assuming three aircraft are coordinating tracking, the system design goal is to achieve full coverage of the target. To this end, the relative angle between each aircraft is set to maintain approximately 360 / N (where N is the number of aircraft) to achieve an optimal distribution. With this configuration, even if the target is obscured in one direction by an obstacle, at least one aircraft will continue to accurately track the target and share its position and pose information with the other aircraft. The other aircraft then quickly adjust their position and pose accordingly to restore coverage of the target, ensuring the continuity and robustness of the tracking mission.

[0105] Step (2) proposes a collaborative mechanism based on adaptive attitude correction for the precise tracking task of multi-rotor aircraft. Multiple aircraft share posture information and tracking status, so that the aircraft that loses the target can adjust its own status according to the posture information of other aircraft and maintain continuous tracking, thereby coping with the problem of target loss that may occur in complex environments.

[0106] The collaborative mechanism based on adaptive attitude correction designed in step (2) is to enable the multi-aircraft system to cope with the problem of target loss in complex environments and maintain continuous tracking.

[0107] The adaptive attitude correction method consists of three core modules: a vision strategy, a controller, and an attitude strategy. At each moment, the system receives observation information and attitude information of each aircraft from the environment and inputs them into the vision strategy, controller, and attitude strategy respectively.

[0108] The vision strategy generates vision-based control commands based on the aircraft's current observations. The controller then outputs the tracking status of each aircraft based on its current state, determining whether it has lost the target. This tracking status, combined with the corresponding aircraft's attitude information, is input into the attitude strategy module, which then outputs attitude-based control commands. Ultimately, each aircraft generates two candidate tracking commands, one derived from the vision strategy and the other from the attitude strategy. The controller's tracking status determines which command will be executed.

[0109] Controller Module Design: The controller module inputs image data processed by a convolutional neural network (CNN) and a long short-term memory network (LSTM). This feature data is then fed into a main network consisting of two fully connected layers with ReLU activation functions for further feature extraction. The feature extraction process of the main network can be expressed mathematically as follows:

[0110] h1=ReLU(W1·x+b1)

[0111] h2=ReLU(W1·h1+b2)

[0112] out=(W3·h2+b3)

[0113] Here, x represents the input feature vector, W1 and W2 are the weight matrices of the fully connected layer, b1 and b2 are bias terms, h1 is the feature vector output by the first layer, h2 is the final feature output after processing by the second layer, and the final output layer network outputs the result out. W3 and b3 are the weight matrix and bias term of the final output layer. The final output shape is (N, 2), where N is the number of aircraft and represents the current tracking status of each aircraft, that is, the probability of target loss.

[0114] For the network output out[i] = [z1, z2], it represents the tracking status of the i-th aircraft, z1 is the original score of the aircraft successfully tracking the target, and z2 is the original score of the aircraft failing to track the target. After the normalized exponential function (Softmax) operation, the probability p is obtained, which is defined as follows:

[0115]

[0116] Where p1 represents the probability of successful tracking of the current aircraft, and p2 represents the probability of the current aircraft losing the target. Here, the Softmax function ensures that p1 + p2 = 1, that is, a probability distribution is generated. Finally, a max operation is performed to determine the index position with the highest probability value in the probability distribution, which can be expressed as:

[0117] index[i]=argmaxp j

[0118] The controller analyzes the current observation images of all aircraft, evaluates whether each aircraft has lost the target, and outputs the corresponding tracking status. index[i] represents the status of the i-th aircraft, 0 indicates target loss, and 1 indicates successful tracking.

[0119] In general, the controller is a binary classification neural network that receives image data and outputs tracking state probabilities, optimized using binary cross-entropy. Its primary responsibility is to determine whether the vehicle should execute the actions generated by the vision policy to continue tracking the target, or the actions generated by the attitude policy to correct its attitude.

[0120] Attitude strategy module design: The module input includes the attitude information of all aircraft and the controller output. To improve the numerical stability and generalization ability of model training, the attitude information needs to be normalized first. Assume that the position information and camera angle of each aircraft are defined as follows:

[0121] Where (x i ,y i ,z i ) represents the position coordinates of the i-th aircraft, and θ i They represent the yaw angle (Yaw) and pitch angle (Pitch) of the aircraft's payload camera.

[0122] d=(d x ,d y ,d z ):d x d y d z is the mean of the initial positions of all aircraft, used for centering. The total number of aircraft is N.

[0123] The normalization of position coordinates can be expressed as:

[0124]

[0125] The angle information is represented by sine and cosine functions to further avoid the discontinuity of periodic angles in model processing:

[0126]

[0127] Among them, σ is the scale factor, which is usually taken as the maximum distance of all aircraft position offsets to scale the data range. Finally, the attitude information and angle information of each aircraft are combined into a feature vector:

[0128] c i =(p″ i ,sin yaw ,cos yaw ,sin pitch ,cos pitch )

[0129] This normalization method not only simplifies the calculation, but also enhances the numerical stability of the data, helping the model focus on the relative posture relationship between aircraft while weakening the impact of absolute position errors.

[0130] Next, the normalized feature vector c of each aircraft is i Combined with the corresponding controller output index[i], it forms the input feature of the posture strategy module:

[0131] gf i =(c i ,index[i])

[0132] These combined features are fed into a gated recurrent unit (GRU) network for temporal information processing. The GRU is a model designed specifically for processing sequential data, effectively capturing dynamic changes along the temporal dimension. In target tracking tasks, the vehicle's controller output data and real-time attitude information are crucial. The controller output reflects whether the vehicle has successfully tracked the target, providing important feedback to the GRU, helping it learn the intrinsic relationship between the vehicle's attitude changes and the target tracking status. This feedback allows the GRU to adjust its hidden state at each time step, dynamically optimizing the decision-making process and enabling the vehicle to respond more accurately in different environments and states.

[0133] The attitude information of all aircraft plays a more important role in this process. The attitude information of the aircraft not only provides the GRU with the historical status of the aircraft (such as the previous posture and tracking status), but also generates globally shared dynamic input by integrating the posture information of other aircraft. This global information sharing mechanism enables intelligent attitude adjustment based on the posture of other aircraft that have successfully tracked the target when an aircraft loses its target. This not only improves the fault tolerance of the system, but also enhances the adaptability of the entire system to complex environments. The working mechanism of the GRU network can be described by the following formula:

[0134] output t =GRU(gf i ,h t-1 )

[0135] where h t-1 is the hidden state at time step t-1, output t is the output vector at time step t. GRU generates a feature sequence Y by dynamically modeling the time series information, which contains the posture, controller output and its time-varying characteristics.

[0136] The feature sequence Y output by the GRU network serves as input to the policy network. Based on this feature information, the policy network generates active tracking control commands for the aircraft. During this process, the policy network comprehensively considers the motion trajectories and attitude changes of all aircraft, as well as the controller's output signals, to output corrective actions for each aircraft. These corrective actions enable the aircraft to adjust its flight path in real time, maintaining its trajectory toward the tracked target, even if the target is out of the current aircraft's field of view. Through this global collaboration and policy adjustment, the system achieves more efficient and stable multi-aircraft collaborative target tracking.

[0137] Generally speaking, the attitude strategy module receives the posture information of all aircraft and outputs corresponding control actions. Its core responsibility is to address the problem of target loss that may occur during the tracking process by integrating the global posture information of multiple drones to generate corrective control instructions. The controller module then determines whether to adopt and execute them based on the current tracking status. Visual Strategy Module Design: The visual strategy mainly makes tracking decisions based on the current image information of the aircraft. First, the observation image of the aircraft is subjected to feature extraction by a convolutional neural network (CNN), which extracts key data containing target motion characteristics from the image. Since target tracking is a time-series task, a long short-term memory network (LSTM) is used to further process the extracted features to capture the dynamic changes of the target in the temporal dimension. Through the combination of CNN and LSTM, the model can obtain rich motion features, which in turn supports subsequent decision-making and outputs accurate tracking action instructions.

[0138] The core responsibility of the visual strategy module is to ensure that the aircraft can always keep the target in the center of the camera's field of view. Based on the image information received by the aircraft, this module generates corresponding visual tracking actions and passes them to the controller module to determine whether to execute them. This part is similar to the paper End-to-end Active Object Tracking via Reinforcement Learning published by Peking University in 2018. (Luo W, Sun P, Zhong F, et al. End-to-end Active Object Tracking via Reinforcement Learning [C]. Proceedings of the 35th International Conference on Machine Learning, 2018: 3286-3295.) Step (3) Based on the global field of view coverage requirement, a collaborative mechanism based on view angle and distance optimization is proposed. The relative distance and view angle distribution between aircraft are included in the optimization range to ensure multi-angle coverage in complex environments, avoid the generation of blind spots, and further enhance the collaborative capability of the multi-aircraft system.

[0139] The main purpose of the collaborative mechanism based on perspective and distance optimization proposed in step (3) is to maintain a reasonable perspective distribution among multiple aircraft. This method incorporates the relative distance and perspective distribution between aircraft into the optimization range to ensure multi-angle coverage in complex environments, avoid the creation of blind spots, and thus enhance the robustness of the entire system.

[0140] In this method, the core implementation method lies in the multiple reward mechanism. Through the design of various reward functions, the aircraft can keep a relatively reasonable distance while tracking the target and maintain the global field of view distribution. Assume that the system consists of N aircraft, and the total reward R obtained by each aircraft at time t is i (t) can be expressed as a combination of two parts:

[0141] R i (t) = α·R local,i (t)+β·R global (t)

[0142] Among them, R local,i (t) represents the independent tracking reward of the i-th aircraft, which mainly measures the tracking accuracy of a single aircraft on the target; R global (t) is the global reward of the system, focusing on the overall perspective and distance optimization goals of the multi-aircraft system; α and β are the weight factors of the local and global rewards, respectively, which are used to adjust the contribution ratio of the two parts of the reward to the total reward.

[0143]

[0144] Here is the independent tracking reward for each aircraft, α(t), β(t) and d(t) represent the horizontal offset, vertical offset and target distance between the target and the center of the field of view, respectively, and α max , β max and d max is the corresponding normalization coefficient, which is used to limit the range of reward values.

[0145] The view distribution reward is designed to encourage the aircraft to track the target from different angles with the target as the center, so as to achieve more comprehensive coverage. Let the view vectors of the i-th and j-th aircraft be and The angle between the two is θ ij To achieve ideal multi-angle coverage, the system defines the desired angle as θ target , then the loss function of view distribution optimization can be expressed as:

[0146]

[0147] where θ target =360 / N, indicating that N aircraft are evenly distributed on a circle centered on the target. This loss function reflects the deviation between the actual distribution and the expected distribution. The loss term is expressed as a negative reward in the reward function and is normalized to [-1, 0], i.e.:

[0148]

[0149] Among them, γangle is the weight factor of the view distribution reward, α angle is the angular error normalization coefficient. The introduction of view distribution optimization enables the system to actively adjust the aircraft's flight direction to ensure tracking from different angles around the target, greatly improving the comprehensiveness of coverage.

[0150] In order to avoid the problem of overlapping perspectives caused by the close distance between aircraft, the algorithm designs a distance loss function to constrain the spatial distribution between aircraft. ij (t)=||P i (t)-P j (t)|| is the distance between the i-th and j-th aircraft, and the expected distance is d target , then the average error between the distance between N aircraft and the expected distance can be expressed as:

[0151]

[0152] The reward function corresponding to the distance loss term is:

[0153]

[0154] Among them, γ distance is the weight factor of the distance reward, α distance is the distance error normalization coefficient. By optimizing the distance between aircraft, we can reduce perspective overlap, improve resource utilization efficiency, and avoid occlusion caused by being too close to the target.

[0155] Finally, the global reward R global (t) is a linear combination of viewing angle and distance rewards:

[0156] R global (t) = R global，angle (t)+R global，distance (t)

[0157] Substituting the reward function designed above into the total reward formula, the aircraft's vision and attitude controllers are trained through a deep reinforcement learning algorithm. This allows them to obtain higher rewards while not only accurately tracking the target but also improving the coverage efficiency of the entire system through a reasonable distribution of viewing angles and distances.

[0158] In the above-mentioned reward function design, when the observation angle distribution or relative distance between aircraft deviates from the ideal state, the perspective distribution reward and distance maintenance reward in the global reward function will be significantly reduced, resulting in a decrease in the total reward value obtained by the decision network. This signal feedback will drive the deep reinforcement learning model to automatically adjust the network parameters, guiding the aircraft to adopt an action strategy that can both maintain target tracking and optimize group collaboration. In actual operation, each aircraft achieves intelligent decision-making through a dual optimization mechanism: on the one hand, it adjusts the camera angle and flight trajectory in real time to accurately stabilize the tracked target in the center of the field of view to maximize the local tracking reward; on the other hand, it continuously evaluates the relative posture relationship with other aircraft, and dynamically optimizes the observation angle and interval distance to obtain a global collaborative reward.

[0159] Example:

[0160] The embodiments of the present invention relate to a multi-rotor aircraft cooperative active target tracking based on adaptive attitude correction and viewing distance optimization, such as Figure 9 As shown, the following steps are included:

[0161] (1) Initialize the virtual 3D environment and set relevant simulation information;

[0162] (2) Construct three modules: posture strategy, controller, and visual strategy, and generate tracking actions based on input information;

[0163] (3) Determine the tracking action adopted by the aircraft based on the controller output and verify the effectiveness of the method.

[0164] In step (1), a virtual 3D environment for training and testing is constructed using Unreal Engine 4, which can simulate the real world with high fidelity. The present invention uses UnrealCV, which provides many convenient APIs for the interaction between reinforcement learning algorithms and virtual environments. Observations are images obtained from the first-person view of the aircraft camera. The actions are discrete and include seven candidate actions: move forward, move backward, turn left, turn right, turn forward and turn left, turn forward and turn right, and stop. The aircraft and the moving target are placed in a spacious closed room, and three aircraft are set to collaborate to complete the active tracking task of the target, such as Figure 2As shown. The aircraft is fixed at a certain altitude and is set to observe the target at a 30-degree angle of view, thereby simulating the working mode of the aircraft in the actual scene. At the beginning of each round of training, the target is randomly initialized in the central area of the room. Correspondingly, the aircraft are evenly distributed around the target, maintaining the set optimal spacing and viewing angle distribution. This optimal distribution initialization setting provides positive feedback for model training, which helps the model to quickly learn the target tracking strategy in the early stages of the mission and improve the stability of the training. The present invention not only evaluates the robustness of the model in a randomly generated environment, but also conducts performance tests in simulated garden and urban scenes, such as Figure 3 These test scenes are closer to the real world, including complex elements such as vegetation, buildings, and irregular obstructions in open layouts, which increases the randomness and challenge of the environment.

[0165] In step (2), at each time step, the aircraft information will be input into three modules, such as Figure 1 As shown in the figure: controller, visual strategy module and attitude strategy module. Specifically, the observation information will be input into the controller and visual strategy module, and the attitude information of all aircraft will be input into the attitude strategy module. The design of the three modules is as follows: Figure 4 As shown. After deep reinforcement learning training, the controller module outputs the tracking status of the current aircraft, that is, whether the target is in the center of the image of the aircraft camera; the visual strategy module outputs the tracking action generated based on the observation information, the goal is to keep the target in the center of the field of view of the aircraft camera; the attitude strategy module outputs the action generated based on the attitude information and tracking status of other aircraft, that is, the posture of the current aircraft is adjusted according to the posture information of the aircraft that successfully tracks the target, so as to maintain continuous tracking of the target. The attitude strategy module implements an adaptive attitude correction mechanism. When the aircraft loses the target due to occlusion or other reasons, it can adjust its own posture and correct its own posture based on the posture information of other aircraft to ensure uninterrupted tracking. At the same time, when generating tracking actions, the attitude strategy module will also optimize the viewing angle and distance, and try to maintain a reasonable distance and viewing angle distribution between multiple aircraft, such as Figure 5 shown.

[0166] In step (3), each aircraft has two sets of tracking actions. The output of the controller module is used to determine which set of tracking actions to use. If the aircraft is still successfully tracking the target, the instructions provided by the vision strategy are executed; if the aircraft has lost the target, the instructions of the attitude strategy are executed first to ensure that the target can be recaptured and tracking can be resumed.

[0167] To verify the effectiveness of the adaptive attitude correction mechanism and view angle and distance optimization design proposed in this paper in the multi-aircraft collaborative active target tracking task, this paper designed a set of comparative experiments using different collaborative tracking methods for performance evaluation. The specific design is as follows:

[0168] 1) Unobstructed environment (baseline): This test environment is set to an ideal environment with no obstructions, and the three aircraft independently perform active tracking tasks.

[0169] 2) Individual Tracking (ST): In this method, the three aircraft independently perform active tracking tasks, without any coordination between them. This setup aims to verify the necessity of inter-vehicle coordination to improve tracking performance.

[0170] 3) Collaborative Tracking (ATC) with Adaptive Attitude Correction: This method introduces an adaptive attitude correction mechanism, allowing three aircraft to collaboratively complete the tracking task. If one aircraft loses the target due to occlusion or other factors, it can adjust its own attitude by obtaining attitude information shared by other aircraft that have successfully tracked the target, thereby restoring the target's tracking capability.

[0171] 4) Collaborative Tracking with Adaptive Attitude Correction and View and Distance Optimization (ATC+VTO): This method incorporates a view and distance optimization strategy in addition to an adaptive attitude correction mechanism. Each aircraft, when performing tracking maneuvers, not only considers the target's position but also comprehensively evaluates its distance and view relationship with other aircraft. This ensures that the entire system maintains good all-round view coverage and a reasonable spatial distribution, further improving tracking stability and efficiency.

[0172] The experimental design tests 100 time steps in 100 random environments, and calculates the average tracking reward and average tracking success rate of the three aircraft. The average tracking reward measures the deviation between the tracked target and the center point of the aircraft camera, normalized to the range of [-1,1]; the average tracking success rate reflects the situation in which each aircraft loses the target when performing the tracking task, and also evaluates the system's adaptive correction mechanism, that is, how to restore tracking with the help of other aircraft after a certain aircraft loses the target. In addition, the test results of the aircraft in an unobstructed random environment are set as the benchmark value, as the most ideal tracking performance indicator, and used to quantify the impact of obstructions on the tracking effect. Since the adaptive posture correction and viewing distance optimization method proposed in the present invention is designed to deal with occlusion problems in complex environments, if the final tracking effect is close to the benchmark value, it can prove the effectiveness of the proposed method.

[0173] Figure 6The following table shows the changing trends in the tracking rewards for the three aircraft using different methods. The experimental results show that the individual tracking (ST) method performs poorly in occluded environments, with the average tracking reward for the three aircraft significantly lower than the baseline. This indicates that occlusions severely impact the individual tracking capabilities of aircraft: when a target is obscured, the aircraft often loses sight of it. Without a coordinated mechanism, reacquiring the target becomes extremely difficult, significantly reducing tracking effectiveness.

[0174] After incorporating the adaptive attitude correction (ATC) mechanism, the average tracking reward of the three aircraft significantly increased. This demonstrates that the attitude correction mechanism effectively mitigates the effects of occlusion. When a vehicle loses its target, it can adjust its view angle through the attitude correction strategy to resume tracking. Furthermore, as long as not all vehicles lose their target simultaneously, at least one vehicle can provide valid attitude information to the others, supporting the proper operation of the correction mechanism. However, due to the lack of global optimization of view angle and distance, the three vehicles may gradually converge during tracking, resulting in view angle overlap, which is detrimental to long-term collaborative tracking. Combining adaptive attitude correction with view angle and distance optimization (ATC+VDO) further improves the average tracking reward of the three vehicles, even approaching the baseline value in an unobstructed environment. This is because the view angle and distance optimization effectively prevents vehicles from getting too close together and overlapping, thereby reducing the likelihood of all vehicles losing their target simultaneously. Compared to the attitude correction mechanism alone, the addition of view angle and distance optimization further enhances the collaborative tracking capability of the multi-vehicle system in occluded environments.

[0175] In summary, the experimental results show that obstructions have a significant impact on the tracking ability of aircraft, but the adaptive attitude correction and view distance optimization mechanism can improve the collaborative tracking ability of aircraft. In particular, under the action of the dual mechanism, the average tracking reward of the three aircraft is close to the benchmark value. The optimal performance of this multi-aircraft collaborative tracking method in complex environments has reached 87.7% of the benchmark value of the unobstructed environment. This shows that the method proposed in this invention is effective in dealing with target occlusion and loss problems in complex environments. In addition, Figure 7 As shown in , the tracking success rate of each method is highly consistent with the average tracking reward results, which further verifies the reliability and applicability of the method. Figure 8 As shown, the 3D scatter plot shows the performance indicators of different methods in the multi-aircraft active target tracking task, including tracking reward, tracking success rate and aircraft distance reward. This result verifies the effectiveness of the perspective and distance optimization mechanism, indicating that reasonable optimization of the distance and perspective distribution of aircraft can not only improve the tracking performance of the system, but also enhance the efficiency and robustness of multi-aircraft collaboration.

[0176] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium is capable of storing a computer program that, when executed by the data processing unit, can execute the invention content of a multi-rotor aircraft collaborative target tracking method based on adaptive posture correction provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0177] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a computer program, i.e., a software product. The computer program software product can be stored in a storage medium and includes a number of instructions for enabling a device including a data processing unit (which can be a personal computer, server, single-chip microcomputer, MCU, or network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0178] The present invention provides a concept and method for cooperative target tracking of multirotor aircraft based on adaptive attitude correction. There are many specific methods and approaches to implement this technical solution. The above is only a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also considered to be within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A multi-rotor aircraft collaborative target tracking method based on adaptive posture correction, characterized in that: The following steps are involved: Step 1: Model the multi-rotor aircraft collaborative target tracking task, modeling it as a precise target tracking task and a global view coverage task, and determine the collaborative relationship and task objectives between the aircraft; Step 2: Design an adaptive attitude correction mechanism to recapture the target based on the attitude information of other aircraft in the precise target tracking task; Step 3: Design a collaborative mechanism based on perspective and distance optimization. Through a weighted combination of local rewards and global rewards, optimize the relative positions and perspective distribution between aircraft to achieve the global perspective coverage task. Step 4: Use multiple aircraft to simultaneously adopt the adaptive attitude correction mechanism and the collaborative mechanism based on view angle and distance optimization to achieve multi-rotor aircraft collaborative target tracking based on adaptive attitude correction.

2. The method for cooperative target tracking of a multi-rotor aircraft based on adaptive attitude correction according to claim 1, characterized in that: Modeling the multirotor cooperative target tracking task described in step 1 includes: Step 1-1: The task of multi-aircraft collaborative active tracking of a single target is divided into the task of precise target tracking and the task of global view coverage, as follows: Precision tracking missions, where when a single aircraft loses its target, other aircraft assist in calibration; Global perception tasks, where a single aircraft considers the distance and view distribution between it and other aircraft when choosing its own actions; In steps 1 and 2, the precise target tracking task and the global view coverage task are modeled respectively.

3. The method for cooperative target tracking of a multi-rotor aircraft based on adaptive attitude correction according to claim 2, characterized in that: The precise target tracking task and the global view coverage task described in steps 1-2 are modeled as follows: Step 1-2-1: Model the precise target tracking task as a multi-agent deep reinforcement learning problem as follows: Assume the input state of the i-th aircraft is s i , contains the position information and attitude angle information of the aircraft, indicating the globally perceived environmental state; Based on the deep reinforcement learning method, the training strategy network π(a|s i ), used to generate action strategies, that is, at each time step t, the strategy network outputs the action a of the i-th aircraft i , used to adjust the attitude and position of the aircraft so that the camera is re-aimed at the target direction; Build a global perception network, rely on the location information provided by other aircraft to infer the possible location of the target, and adjust the aircraft's own posture and position accordingly, so that the camera is aimed at the potential target area; Step 1-2-2 covers the global perspective task as follows: Under the coordinated tracking of multiple aircraft, the mission goal is to achieve all-round coverage of the target.

4. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 3, characterized in that: Design an adaptive posture correction mechanism as described in step 2, including: Vision strategy module, controller module and posture strategy module; At each moment, the observation information and attitude information of each aircraft are input into the visual strategy module, controller module and attitude strategy module respectively; The self-vision strategy module and the self-attitude strategy module generate candidate tracking instructions respectively, which are executed by the controller module. If the aircraft does not lose the target, the candidate tracking instructions generated by the vision strategy module are executed; otherwise, the candidate tracking instructions generated by the attitude strategy module are executed.

5. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 4, characterized in that: The controller module includes: The observed image features processed by the convolutional neural network (CNN) and the long short-term memory (LSTM) network are received and input into the first network consisting of two fully connected layers for state classification, which is represented as follows: h1=ReLU(W1·x+b1) h2=ReLU(W2·h1+b2) out=(W3·h2+b3) Among them, x represents the input observation image feature vector, W1 and W2 are the weight matrices of the fully connected layer, b1 and b2 are bias terms, ReLU represents the activation function, W3 and b3 represent the weight matrix and bias term of the final output layer, and the final output of the first network is out, which is a vector of shape [N, 2], denoted as: out[i]=[z1,z2] Among them, out[i] is the status output of the i-th aircraft, z1 corresponds to the original score of aircraft i successfully tracking the target, and z2 corresponds to the original score of aircraft losing the target; After the normalized exponential function Softmax operation, the probability P is obtained p , which is expressed as follows: Where P1 represents the probability of successfully tracking the target, P2 represents the probability of losing the target, P1+P2=1; Through the maximum value operation, it is expressed as follows: index[i]=arg max P p Get the classification result index[i] of the current state, that is, complete the judgment of whether the i-th aircraft has lost the target.

6. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 5, characterized in that: The posture strategy module includes: Assume the position information of the i-th aircraft and the camera angle p i Expressed as: Among them, (x i ,y i ,z i ) is the position coordinate of the aircraft, and θ i Respectively represent the yaw angle and pitch angle of the camera on the aircraft; Normalize the position information of the aircraft and the camera angle to obtain the normalized pose information combined feature vector c of the i-th aircraft i , which is expressed as follows: c i =(p″ i ,without yaw ,cos yaw ,without pitch ,cos pitch ) Among them, p″ i is the normalized position information of the aircraft, sin yaw ,cos yaw ,sin pitch ,cos pitch is the normalized angle information, which are expressed as follows: Among them, (d x ,d y ,d z ) is the mean of the initial positions of all aircraft, σ is the scale factor, i.e. the maximum distance of the position offset of all aircraft; Combined with the controller's output index[i], the input feature gf of the posture strategy module is obtained i , which is expressed as follows: gf i =(c i ,index[i]) The input feature gf of the i-th aircraft i The input is fed into the GRU network for temporal information processing, which is expressed as follows: output t =GRU(gf i ,h t-1 ) Among them, h t-1 is the hidden state at time step t-1, output t is the output vector at time step t; Output vector output t The input is sent to the policy network, which uses the Actor-Critic structure to learn and optimize the drone control strategy and generate correction instructions as candidate tracking instructions output by the posture policy module.

7. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 6, characterized in that: The visual strategy module includes: The observation image of the aircraft is subjected to feature extraction by the convolutional neural network (CNN) to obtain the target motion features; Use the long short-term memory network (LSTM) to capture the dynamic changes of the target in the time dimension and obtain the time change features; The target motion features and time change features are combined and input into the visual strategy module, which outputs the candidate vision-based tracking instructions for the aircraft.

8. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 7, characterized in that: The design described in step 3 is based on a synergistic mechanism of view and distance optimization, including: Assume that the total reward obtained by the i-th aircraft at time t is R i (t), is expressed as follows: R i (t)=α·R local，i (t)+β·R global (t) Among them, R local,i (t) represents the independent tracking reward of the i-th aircraft, i.e., the local reward, R global (t) is the global reward, α and β are the weight factors of local and global rewards respectively; The total reward is used to adjust network parameters for deep reinforcement learning methods in precise target tracking tasks.

9. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 8, characterized in that: The local reward R described in step 3 local,i (t), is calculated as follows: Among them, α(t), β(t) and d(t) represent the horizontal offset, vertical offset and target distance between the target and the center of the aircraft’s field of view, respectively. max , β max and d max is the corresponding normalization coefficient.

10. The method for cooperative target tracking of multi-rotor aircraft based on adaptive attitude correction according to claim 9, characterized in that: The global reward R described in step 3 global (t) is a linear combination of the viewing angle reward and the distance reward, expressed as follows: R global (t)=R global，angle (t)+R global，distance (t) Among them, the perspective reward R global，angle (t), is calculated as follows: Among them, γ angle is the weight factor of the view distribution reward, α angle is the angle error normalization coefficient, L angle is the view distribution optimization loss function, expressed as: Where the desired angle is θ target , the angle between the perspectives of the i-th and j-th aircraft is θ ij ; Distance Reward R global，distance (t), is calculated as follows: Among them, γ distance is the weight factor of the distance reward, α distance is the distance error normalization coefficient, the distance loss function L distance , expressed as: The expected distance is d target , the distance between the i-th and j-th aircraft is d ij (t).

Images