Unmanned ship trajectory prediction method based on target driving and dynamic constraint

Abstract

The invention relates to an unmanned ship trajectory prediction method based on target driving and dynamic constraint. The method comprises the following steps: firstly, generating a possible future target position of the unmanned surface vehicle by utilizing a target generation network based on U-Net and combining a historical track and ocean scene information of the unmanned surface vehicle; and then, based on a kinetic equation, carrying out physical constraint on the trajectory to ensure that the generated trajectory is smooth and conforms to the physical motion law of the unmanned ship. Meanwhile, in order to simulate random disturbance effects in marine environments such as water flow and wind power, dynamic random modeling is further introduced, so that the practical adaptability and robustness of trajectory prediction are enhanced. Finally, through joint optimization of a target loss function of a target generation network, an L2 loss function generated by a dynamic trajectory and a distribution modeling loss function of a dynamic random item, the precision and stability of trajectory prediction are comprehensively improved.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned boat intelligent navigation and path prediction and planning, and is particularly aimed at the needs of unmanned boat trajectory prediction in complex marine environments and high-dynamic mission scenarios. Background Art

[0002] Unmanned boat trajectory prediction is one of the core technologies of intelligent navigation and mission planning of unmanned boats, which directly determines the navigation accuracy and mission execution efficiency of unmanned boats in complex marine environments. However, limited by the technical bottlenecks of traditional methods, the current trajectory prediction technology still has many shortcomings in terms of dynamics and adaptability. For example, traditional trajectory prediction methods usually rely on fixed rules or simple prediction models, which makes it difficult to comprehensively consider the target driving characteristics, dynamic constraints and random disturbances in the motion of unmanned boats in complex environments. This results in the generated trajectory often not being physically feasible in actual operation and unable to adapt to the complex and changeable marine environment, thereby reducing the mission execution efficiency and stability of the unmanned boat.

[0003] In addition, the marine environment has significant dynamic characteristics, such as water flow, wind and sudden environmental changes, which pose great challenges to the trajectory planning of unmanned boats. However, existing methods often ignore the impact of these dynamic disturbances and find it difficult to accurately model the actual motion trajectory. This technical limitation directly affects the performance of unmanned boats in tasks such as maritime cruising, search and rescue, and target tracking. It not only limits the application scenarios of unmanned boat technology, but also affects the safety and reliability of unmanned boats in highly dynamic environments.

[0004] In order to solve the above problems, the present invention proposes a method for unmanned boat trajectory prediction based on target drive and dynamic constraints. By combining the data fitting ability of the deep learning model with the physical constraints of the dynamic equation, this method can generate a smooth trajectory that conforms to physical laws, and effectively respond to random disturbances in complex environments through dynamic randomness modeling, thereby significantly improving the accuracy and adaptability of unmanned boat trajectory prediction. This method provides intelligent support for the mission planning and execution of unmanned boats in complex marine environments, greatly expanding the application scope of unmanned boat technology, and has important research value and application prospects. Summary of the invention

[0005] The present invention proposes a method for unmanned boat trajectory prediction based on target drive and dynamic constraints, aiming to solve the problems of inaccurate trajectory prediction, uneven path and insufficient adaptability to environmental interference in traditional methods in complex environments. This method combines ocean scene information and historical trajectory of unmanned boats through a target generation network to accurately predict future target positions, and uses dynamic equations to physically constrain the motion trajectory to ensure trajectory smoothness and feasibility. In addition, the method further introduces dynamic randomness modeling to simulate the random disturbance effects in the ocean environment such as water flow and wind, and enhance the realistic adaptability and robustness of trajectory prediction. This technology can be widely used in mission scenarios such as ocean cruising, target tracking, search and rescue support for unmanned boats, providing efficient and intelligent path planning and dynamic decision support for unmanned boats, significantly improving their mission execution capabilities and reliability in complex ocean environments.

[0006] The purpose of the present invention is to provide an unmanned boat prediction method based on target drive and dynamic constraints to solve the problems raised in the above background technology. The present invention provides the following technical solutions:

[0007] A target-driven and dynamic constraint-based unmanned vehicle trajectory prediction method combines the data fitting ability of the deep learning model with the physical constraints of the dynamic equation, aiming to improve the accuracy and reliability of trajectory prediction and tracking control of unmanned vehicles in complex marine environments. The overall process of the method is as follows: First, in view of the characteristics that unmanned vehicles usually have a clear destination when sailing or performing tasks in the marine environment, a target generation network based on U-net is used to generate possible target positions by combining marine scene information and the historical trajectory of the unmanned vehicle. Subsequently, according to the generated target position, the dynamic equation is introduced to constrain the movement mode of the unmanned vehicle at each time step to ensure that the generated trajectory is both physically feasible and smooth. Finally, in order to simulate the disturbance effect of the marine environment (such as water flow, wind, etc.) on the unmanned vehicle, dynamic random terms are further added to the dynamic equation to enhance the modeling ability of the motion trajectory in the real environment. Through the combination of target generation, dynamic constraints and random disturbances, this method significantly improves the understanding and prediction ability of the movement mode of unmanned vehicles in complex marine environments, and provides favorable support for the path planning and control of unmanned vehicles in practical application scenarios. The specific steps are as follows:

[0008] Step A: Based on the U-net unmanned boat target generation network, the trajectory information of the unmanned boat is aligned with the scene information in the form of an image. The U-net structure is used to efficiently integrate global information and local detail information, and the historical movement pattern of the unmanned boat is comprehensively learned. In combination with the scene segmentation results, the unmanned boat's possible future targets are accurately predicted, including the following steps:

[0009] Step A1: The historical trajectory of the unmanned boat is spliced ​​with the scene information as the input of the target generation network.

[0010] The task of unmanned boat trajectory prediction is defined as providing the unmanned boat in the historical t p The set of positions at each time step and ocean scene diagram I, predicting the unmanned boat in the future f The position of the time step

[0011] The historical trajectory of unmanned boats i Convert to trajectory heatmap The specific conversion formula is as follows:

[0012]

[0013] Among them, (i, j) represents the pixel coordinates, o t =(x t ,y t ) is the pixel coordinate of the unmanned boat at time t. p represents the number of historical time steps, and H×W is the size of the heat map.

[0014] The ocean scene image I is segmented into segmentation maps through an image segmentation network (such as U-net) Used to characterize the current marine environment properties, N c is the number of output categories of the segmentation network. For example, the segmentation map can indicate which areas are feasible and which areas contain dangerous objects such as reefs.

[0015] The trajectory heat map and segmentation map are spliced ​​together as the input feature H of the target generation network in .

[0016] Step A2: The target generation network adopts the encoder-decoder structure based on U-net and accepts the H generated in step A. in And output the prediction of the future target distribution of the unmanned boat.

[0017] In the encoding stage, the input feature H in Through six encoding blocks consisting of maximum pooling and convolution, the spatial resolution is gradually halved and the number of channels is doubled, ensuring that the network can capture deeper features while retaining key local detail information and global semantic information. The final encoder output H m ={H 1 ,H 2, …,H 6}, which contains deep feature representation and intermediate features of multiple scales.

[0018] In the decoding stage, the decoder uses H mThe deep feature representation in is taken as input. The decoding process consists of multiple decoding blocks. In each decoding block, the spatial resolution is first gradually improved through bilinear interpolation operations, and then the feature representation is further optimized through convolution operations to extract richer semantic information. This process is performed layer by layer until the resolution of the feature map is restored to the same size as the original input. In order to further enhance the feature reconstruction capability, the decoding stage fuses the intermediate features from the corresponding layer of the encoder through jump connections at each layer, thereby integrating multi-scale local and global information and retaining the important details and semantic features extracted during the encoding process. The high-resolution feature H finally output by the decoder u , the mathematical expression of the decoding process is as follows:

[0019]

[0020] Among them, Conv represents the convolution operation, represents a bilinear upsampling operation, Represents a concatenation operation.

[0021] H u After being processed by the Sigmoid activation function pixel by pixel, the future t of the unmanned boat is generated. f The position probability distribution at time t represents the probability that the unmanned boat will appear at the position (x, y) in the future time step t, which is recorded as The last moment corresponds to the target probability distribution P(x,y,t p +t f ). By combining the multi-scale features of the encoder and decoder, the target generation network can provide accurate and reliable predictions for the distribution of future targets of the unmanned boat.

[0022] Step A3: Generate the target probability distribution P(x, y, t) output by the target generation network in step A3 p +t f ), sampling generates predictions of K possible future targets.

[0023] First, for P(x,y,t p +t f ) Apply the Soft-argmax function and select the point with the largest probability value as the first target point G a .

[0024] Then, from P(x,y,t p +t f ) and filter out those with probability lower than 0.01×max(P(x,y,t p +tf point.

[0025] Apply the clustering algorithm to the remaining candidate points, divide them into K-1 clusters, and select the center point of each cluster as the target point.

[0026] Finally, combined with G a Together with these cluster center points, the target point set G = {G a}∪{G 1 ,G 2 ,…,G K-1}, representing the K possible future target points of the unmanned boat.

[0027] Step B: By building a dynamic model of the unmanned boat, its motion characteristics (including speed, acceleration, heading angle, etc.) are coupled with the future target point. Based on the constraints of the dynamic equation, the state information of the unmanned boat is updated in each time step to ensure that the trajectory generation process conforms to physical feasibility and avoids the generation of unreasonable trajectories. Specifically, the following steps are included:

[0028] Step B1: Study the dynamic equations that control the motion of the unmanned boat.

[0029] The unmanned boat first determines a future goal and moves based on the goal. The state of the unmanned boat at a certain time t is set to Where o(t) represents the current position at time t, is the vector velocity (i.e. the first-order derivative of o(t)), which can be calculated by the finite difference method. At a certain time T in the future, given the target state of the unmanned boat is q(T), its motion can be expressed as the following equation:

[0030]

[0031] Among them, q(0) is the initial state of the unmanned boat; f θ is the control function, which describes the motion law of the unmanned boat and is determined by the current time t, the current state q(t) and the target state q(T); α t (t,q t:t-M ) is dynamic randomness, simulating the small disturbance of the ocean environment to the unmanned boat, and depends on the historical state of the past M time steps.

[0032] This equation can accurately model the motion behavior of the unmanned boat, thereby predicting the state q(t) of the unmanned boat at any time in the time interval [0,T]. Convert the above continuous time motion equation into discrete time form. Since o(t) has second-order differentiability, the Taylor series expansion of q(t) can obtain the following formula:

[0033]

[0034] Where Δt is the discretized time interval, represents the acceleration, which is determined by the driving force of the unmanned boat to the future target. It is a deterministic term, so the acceleration term can be effectively learned by observing the state changes of the unmanned boat at each discrete time interval and the dynamic randomness term α t (t,q t:t-M ), thereby achieving accurate trajectory prediction.

[0035] Step B2: Study the acceleration term of the unmanned boat That is, learn the target driving force that the unmanned boat is subjected to at each moment.

[0036] The motion trend of the unmanned boat is constrained by its current position and future goal, which is reflected in the changes of speed and heading angle, that is, At time t, the desired navigation direction of the unmanned boat is e t Determined by the target position o(T) and the current position o(t), Therefore, without environmental interference, the unmanned boat will Adjust to desired speed

[0037] In order to simulate the speed change of the unmanned boat when approaching the target, it is necessary to dynamically update the current speed at each time step, so the definition Therefore, the calculation formula for the expected speed at each time step is changed to

[0038] therefore, The item indicates that the pedestrian will change the current speed Change to desired speed The tendency is defined as:

[0039]

[0040] Where τ represents the time required to adjust the current speed to the desired speed, which is determined by the neural network Learned.

[0041] Step B3: Study the dynamic random term α of the unmanned boat t (t,q t:t-M ), simulating the environmental disturbance to the unmanned boat.

[0042] After determining the future goal of the unmanned boat and using the dynamic equation to constrain the movement of the unmanned boat at each time step, the future trajectory can be predicted. However, the unmanned boat will be affected by complex interference in the marine environment (such as water flow, wind, etc.) during actual navigation, which may cause deviations between the actual navigation trajectory and the predicted trajectory. To this end, it is necessary to model the dynamic randomness term α t (t,q t:t-M) to reflect the impact of environmental interference on the movement of unmanned boats.

[0043] Given the predicted position without considering interference and the actual observation position o t , dynamic randomness α t Defined as the error between:

[0044]

[0045] In order to accurately model α t The distribution and its time-varying characteristics are assumed to depend on the short-term historical trajectory q of the unmanned boat. t:t-M , the historical trajectory implicitly contains the environmental factors and the motion pattern information of the unmanned boat. Based on this assumption, the dynamic randomness α t The conditional probability distribution of can be expressed as:

[0046] P(α t |q t:t-M )=∫P(α t |q t:t-M ,z)P(z)dz

[0047] Among them, z represents the latent variable, which is used to capture α t distribution shape, P(z) is the prior distribution of z.

[0048] In order to t To model, assume that there is a mapping Q(z|α t ,q t:t-M ), which represents the latent variable z under a given error α t and historical trajectory q t:t-M The posterior distribution at the time, and z follows a normal distribution in the latent space. By minimizing the KL divergence between the variational posterior distribution Q and the true distribution P, the dynamic randomness α can be learned t The conditional distribution of , thus enabling the model to model the deviations caused by environmental perturbations.

[0049] Step G: The network training of this method mainly consists of the training of the target generation network, the training of the dynamic equation constraint network and the dynamic random term α t (t,q t:t-M ) training consists of three parts:

[0050] Step G1: Training of target generation network. The task of the target generation network is to generate the target corresponding to the unmanned boat at time step t. p +t f The actual position of the target is as close as possible. The target probability distribution predicted by the network output is The actual position of the unmanned boat is converted into aH Gaussian heatmap of P. The target generation network is trained using binary cross entropy loss Optimize, the specific formula is:

[0051]

[0052] Step G2: Training of the dynamics equation constraint network. The dynamics equation constraint network aims to minimize the predicted position at each time step. and the real location t To ensure the stability of training, in each iteration, we assume that the first M+1 frames of the trajectory are known, and gradually predict the remaining trajectory through forward propagation. Using L2 loss Training network, the specific formula is:

[0053]

[0054] Step G3: Training of dynamic random term. Dynamic random term α φ (t,q t:t-M ) is used to fit the random distribution caused by environmental interference. As above, in each training iteration, it is assumed that the first M+1 frames of the trajectory are known, and the subsequent trajectory is predicted step by step, and Loss training, the specific formula is:

[0055]

[0056] The first term is the reconstruction loss, which measures the randomness of the prediction. and true randomness α t The second term is the KL divergence loss, which is used to regularize the potential distribution z to ensure the interpretability and learning effect of the random distribution, and λ is the weight parameter.

[0057] Step G4: Overall objective function is the target generation loss Trajectory Loss and dynamic random term loss The specific formula is:

[0058]

[0059] Among them, λ 1 and λ 2 is a weight parameter used to balance the contributions of target generation, trajectory generation, and randomness modeling.

[0060] Step H: The unmanned boat trajectory prediction method based on target drive and dynamic constraints is mainly used to improve the trajectory planning and prediction capabilities of unmanned boats in complex marine environments. In practical applications, unmanned boats often need to achieve efficient task execution through accurate trajectory prediction. For example, during cruising, the possible target position of the unmanned boat is predicted by combining the historical trajectory and scene information through the target generation network, and the dynamic equation is used to constrain its motion trajectory to simulate the impact of environmental interference such as water flow and wind on the motion, so as to more realistically reflect the actual trajectory changes of the unmanned boat in the dynamic marine environment. This trajectory prediction capability can not only assist the unmanned boat in optimizing path planning, but also support its real-time decision-making and path adjustment in search and rescue, cruising, law enforcement and other scenarios, significantly improving the efficiency and reliability of unmanned boats in complex tasks, and providing strong technical support for the realization of intelligent autonomous navigation.

[0061] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0062] The present invention proposes a comprehensive framework that can accurately predict the trajectory of unmanned boats by combining the target-driven trajectory generation method and the physical constraints of the dynamic equation, overcoming the limitations of the traditional method in complex marine environments, such as uneven trajectory generation, physical impracticability, and poor environmental adaptability. Specifically, the present invention generates the possible target position of the unmanned boat through a target generation network based on U-net, combined with ocean scene information and historical trajectories; then introduces the dynamic equation to constrain the trajectory to ensure that the trajectory generation conforms to physical laws and has smoothness; at the same time, for complex disturbance factors such as water flow and wind, dynamic randomness modeling is further added to improve the adaptability of trajectory prediction to the real environment. Through multi-level optimization of target generation, dynamic constraints and dynamic disturbance modeling, the present invention significantly improves the trajectory prediction accuracy and motion planning ability of unmanned boats in complex marine environments, and provides reliable technical support for the efficient execution of unmanned boats in tasks such as ocean cruising, search and rescue, and target tracking, which has important research and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0064] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The core task of unmanned boat trajectory prediction is to predict the possible position of the unmanned boat in the future based on the position information of the unmanned boat in the past period of time and the current scene environment map. The unmanned boat trajectory prediction method based on target drive and dynamic constraints in this embodiment aims to solve the problems of non-smooth trajectory prediction, physical impracticability and insufficient adaptability to environmental disturbances in complex marine environments. First, the possible future target position of the unmanned boat is generated by combining the historical trajectory and marine scene information of the unmanned boat using the target generation network based on U-Net. Then, according to the result of target generation, the motion trajectory of the unmanned boat is constrained by the dynamic equation to ensure that the generated trajectory is both in line with the laws of physical motion and smooth. Subsequently, dynamic randomness modeling is introduced for random interference effects in marine environments such as water flow and wind, further improving the adaptability and robustness of trajectory prediction to complex environments. Finally, the entire prediction model is trained and performance improved through the joint optimization of the target loss of the target generation network, the L2 loss function of the dynamic trajectory generation, and the distribution modeling loss of the dynamic random term. This embodiment can accurately predict the movement trajectory of unmanned boats in complex marine environments by jointly optimizing the trained trajectory prediction model, providing reliable technical support and broad application prospects for path planning and dynamic decision-making of unmanned boats in cruising, search and rescue, and law enforcement missions.

[0067] See also Figure 1 The process of the unmanned boat trajectory prediction method based on target drive and dynamic constraints of the present invention is as follows: Figure 1 As shown, the specific steps are as follows:

[0068] The task of unmanned boat trajectory prediction is defined as providing the unmanned boat in the past t p The set of positions at each time step and scene environment map I, predicting the unmanned boat in the future t f The position of the time step

[0069] First, the historical trajectory of unmanned boats i Convert to trajectory heatmap The specific conversion formula is as follows:

[0070]

[0071] Among them, (i, j) represents the pixel coordinates, o t =(x t ,y t ) is the pixel coordinate of the unmanned boat at time t.

[0072] Then, the scene graph I is passed through an image segmentation network (such as U-net) to generate a segmentation map Used to characterize the specific attributes of the current marine environment (N c is the total number of classes output by the segmentation network). For example, the segmentation map can indicate which areas are feasible and which areas contain dangerous objects such as reefs. Finally, the concatenated trajectory heat map and segmentation map are used as the input feature H of the target generation network in .

[0073] After getting the input feature H in The target generation network adopts the encoder-decoder structure based on U-net to accept H in And output the prediction of the future target distribution of the unmanned boat. In the encoding stage, the input feature H in Through six encoding blocks consisting of maximum pooling and convolution, the spatial resolution is gradually halved and the number of channels is doubled, ensuring that the network can capture deeper features while retaining key local detail information and global semantic information. The final encoder output H m ={H 1 ,H 2, …,H 6}, including deep feature representation and intermediate features of multiple scales. In the decoding stage, the decoder uses H m The deep feature representation in is taken as input. The decoding process consists of multiple decoding blocks. In each decoding block, the spatial resolution is first gradually improved through bilinear interpolation operations, and then the feature representation is further optimized through convolution operations to extract richer semantic information. This process is performed layer by layer until the resolution of the feature map is restored to the same size as the original input. In order to further enhance the feature reconstruction capability, the decoding stage fuses the intermediate features from the corresponding layer of the encoder through jump connections at each layer, thereby integrating multi-scale local and global information and retaining the important details and semantic features extracted during the encoding process. The high-resolution feature H finally output by the decoder u , the mathematical expression of the decoding process is as follows:

[0074]

[0075] Among them, Conv represents the convolution operation, represents a bilinear upsampling operation, Represents a concatenation operation.

[0076] H uAfter being processed by the Sigmoid activation function pixel by pixel, the future t of the unmanned boat is generated. f The position probability distribution at time t represents the probability that the unmanned boat will appear at the position (x, y) in the future time step t, which is recorded as The last moment corresponds to the target probability distribution P(x,y,t p +t f ). By combining the multi-scale features of the encoder and decoder, the target generation network can provide accurate and reliable predictions for the distribution of future targets of the unmanned boat.

[0077] After obtaining the target probability distribution P(x,y,t p +t f ), sample and generate predictions of K possible future targets. First, for P(x,y,t p +t f ) Apply the Soft-argmax function and select the point with the largest probability value as the first target point G a Then, from P(x,y,t p +t f ) and filter out those with probability lower than 0.01×max(P(x,y,t p +t f )) points. Apply the clustering algorithm to the remaining candidate points, divide them into K-1 clusters, and select the center point of each cluster as the target point. Finally, combine G a Together with these cluster center points, the target point set G = {G a}∪{G 1 ,G 2 ,…,G K-1}, representing the K possible future target points of the unmanned boat.

[0078] Next, the dynamic equations that control the motion of the unmanned boat are studied to constrain the motion mode of the unmanned boat.

[0079] The unmanned boat first determines a future goal and moves based on the goal. The state of the unmanned boat at a certain time t is set to Where o(t) represents the current position at time t, is the vector velocity (i.e. the first-order derivative of o(t)), which can be calculated by the finite difference method. At a certain time T in the future, given the target state of the unmanned boat is q(T), its motion can be expressed as the following equation:

[0080]

[0081] Among them, q(0) is the initial state of the unmanned boat; f θis the control function, which describes the motion law of the unmanned boat and is determined by the current time t, the current state q(t) and the target state q(T); α t (t,q t:t-M ) is dynamic randomness, simulating the small disturbance of the ocean environment to the unmanned boat, and depends on the historical state of the past M time steps.

[0082] This equation can accurately model the motion behavior of the unmanned boat, thereby predicting the state q(t) of the unmanned boat at any time in the time interval [0,T]. Convert the above continuous time motion equation into discrete time form. Since o(t) has second-order differentiability, the Taylor series expansion of q(t) can obtain the following formula:

[0083]

[0084] Where Δt is the discretized time interval, represents the acceleration, which is determined by the driving force of the unmanned boat to the future target. It is a deterministic term, so the acceleration term can be effectively learned by observing the state changes of the unmanned boat at each discrete time interval and the dynamic randomness term α t (t,q t:t-M ), thereby achieving accurate trajectory prediction.

[0085] (1) Study the acceleration term of the unmanned boat That is, learn the target driving force that the unmanned boat is subjected to at each moment.

[0086] The motion trend of the unmanned boat is constrained by its current position and future goal, which is reflected in the changes of speed and heading angle, that is, At time t, the desired navigation direction of the unmanned boat is e t Determined by the target position o(T) and the current position o(t), Therefore, without environmental interference, the unmanned boat will Adjust to desired speed

[0087] In order to simulate the speed change of the unmanned boat when approaching the target, it is necessary to dynamically update the current speed at each time step, so the definition Therefore, the calculation formula for the expected speed at each time step is changed to

[0088] therefore, The item indicates that the pedestrian will change the current speed Change to desired speed The tendency is defined as:

[0089]

[0090] Where τ represents the time required to adjust the current speed to the desired speed, which is determined by the neural network Learned.

[0091] (2) Study the dynamic random term α of the unmanned boat t (t,q t:t-M ), simulating the environmental disturbance to the unmanned boat.

[0092] After determining the future goal of the unmanned boat and using the dynamic equation to constrain the movement of the unmanned boat at each time step, the future trajectory can be predicted. However, the unmanned boat will be affected by complex interference in the marine environment (such as water flow, wind, etc.) during actual navigation, which may cause deviations between the actual navigation trajectory and the predicted trajectory. To this end, it is necessary to model the dynamic randomness term α t (t,q t:t-M ) to reflect the impact of environmental interference on the motion of the unmanned boat. Given the predicted position without considering interference and the actual observation position o t , dynamic randomness α t Defined as the error between:

[0093]

[0094] In order to accurately model α t The distribution and its time-varying characteristics are assumed to depend on the short-term historical trajectory q of the unmanned boat. t:t-M , the historical trajectory implicitly contains the environmental factors and the motion pattern information of the unmanned boat. Based on this assumption, the dynamic randomness α t The conditional probability distribution of can be expressed as:

[0095]

[0096] Among them, z represents the latent variable, which is used to capture α t distribution shape, P(z) is the prior distribution of z.

[0097] In order to t To model, assume that there is a mapping Q(z|α t ,q t:t-M ), which represents the latent variable z under a given error α t and historical trajectory q t:t-M The posterior distribution at the time, and z follows a normal distribution in the latent space. By minimizing the KL divergence between the variational posterior distribution Q and the true distribution P, the dynamic randomness α can be learned t The conditional distribution of , thus enabling the model to model the deviations caused by environmental perturbations.

[0098] The network training of this method mainly consists of the training of the target generation network, the training of the dynamic equation constraint network and the dynamic random term α φ (t,q t:t-M ) training consists of three parts:

[0099] (1) Training of the target generation network. The task of the target generation network is to generate p +t f The actual position of the target is as close as possible. The target probability distribution predicted by the network output is The actual position of the unmanned boat is converted into a H Gaussian heatmap of P. The target generation network is trained using binary cross entropy loss Optimize, the specific formula is:

[0100]

[0101] (2) Training of the dynamic equation constraint network. The dynamic equation constraint network aims to minimize the predicted position at each time step. and the real location t To ensure the stability of training, in each iteration, we assume that the first M+1 frames of the trajectory are known, and gradually predict the remaining trajectory through forward propagation. Using L2 loss Training network, the specific formula is:

[0102]

[0103] (3) Training of dynamic random terms. Dynamic random term α t (t,q t:t-M ) is used to fit the random distribution caused by environmental interference. As above, in each training iteration, it is assumed that the first M+1 frames of the trajectory are known, and the subsequent trajectory is predicted step by step, and Loss training, the specific formula is:

[0104]

[0105] The first term is the reconstruction loss, which measures the randomness of the prediction. and true randomness α t The second term is the KL divergence loss, which is used to regularize the potential distribution z to ensure the interpretability and learning effect of the random distribution, and λ is the weight parameter.

[0106] Overall objective function is the target generation loss Trajectory loss and dynamic random term loss The specific formula is:

[0107]

[0108] Among them, λ 1 and λ 2 is a weight parameter used to balance the contributions of target generation, trajectory generation, and randomness modeling.

[0109] The unmanned boat trajectory prediction method based on target drive and dynamic constraints is mainly used to improve the trajectory planning and prediction capabilities of unmanned boats in complex marine environments. In practical applications, unmanned boats often need to achieve efficient task execution through accurate trajectory prediction. For example, during cruising, the possible target position of the unmanned boat is predicted by combining the historical trajectory and scene information through the target generation network, and its motion trajectory is constrained by the dynamic equation, simulating the impact of environmental interference such as water flow and wind on the motion, so as to more realistically reflect the actual trajectory changes of the unmanned boat in the dynamic marine environment. This trajectory prediction capability can not only assist the unmanned boat in optimizing path planning, but also support its real-time decision-making and path adjustment in search and rescue, cruising, law enforcement and other scenarios, significantly improving the efficiency and reliability of unmanned boats in complex tasks, and providing strong technical support for the realization of intelligent autonomous navigation.

[0110] The present invention proposes a method for unmanned boat trajectory prediction based on target drive and dynamic constraints by combining target generation network, dynamic constraints and dynamic randomness modeling, and provides an effective solution to the problems of non-smooth trajectory prediction, physical impracticability and insufficient adaptability to environmental interference in complex marine environments. Specifically, the present invention accurately predicts the possible target position of the unmanned boat through the target generation network combined with marine scene information and historical trajectories; then physically constrains the trajectory based on the dynamic equation to ensure that the generated trajectory is smooth and meets physical feasibility; at the same time, dynamic randomness modeling is used to simulate random interference such as water flow and wind in the marine environment, further improving the adaptability and robustness of trajectory prediction. This method overcomes the limitation of traditional trajectory prediction methods that it is difficult to accurately model dynamic changes in complex environments, significantly improves the accuracy and stability of unmanned boat trajectory prediction, provides reliable technical support for marine cruises, search and rescue and law enforcement tasks, and greatly enhances the mission execution capability and application scope of unmanned boats in complex marine environments.

[0111] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the trajectory of an unmanned vehicle based on target drive and dynamic constraints, characterized in that: The steps include: Step S1. Future target generation phase: The historical trajectory data X of the unmanned boat i Convert to trajectory heatmap The ocean scene environment map I is generated into a segmentation map through the image segmentation network The trajectory heat map H t and segmentation map H s Concatenation, as the input feature H of the target generation network in , where t p Represents the number of historical time steps, N c is the number of output types of the segmentation network, H×W is the size of the heat map; the target generation network based on U-net inputs the input feature H in , generate the probability distribution P(x,y,t p +t f ), and sample and generate K possible future target point sets G; Step S2. Dynamic constraint stage: Based on the future target point set G generated in step S1, the dynamic equation of the unmanned boat is constructed to physically constrain the motion trajectory of the unmanned boat; Step S3. Trajectory optimization stage: jointly train the target generation network, the dynamic equation constraint network and the dynamic random term network, and generate the loss function of the target generation network through the target The dynamic equation constrains the network's trajectory loss And the dynamic random term loss function of the dynamic random term network Weighted combination to optimize the overall objective function Generate smooth trajectories that conform to physical laws and adapt to environmental disturbances.

2. The unmanned boat trajectory prediction method based on target drive and dynamic constraints according to claim 1 is characterized in that: The specific execution steps of the target generation network of U-net in step S1. are as follows: The encoding stage is used to receive the input feature H in , and the input features are processed step by step through six encoding blocks, each of which contains a maximum pooling layer and a convolutional layer to gradually halve the spatial resolution and double the number of channels, outputting H m ={H1,H 2, …,H6}, including deep feature representation and intermediate features at multiple scales; A decoding stage is used to receive the deep feature representation, and gradually restore the spatial resolution and optimize the feature representation through multiple decoding blocks. Each decoding block includes a bilinear interpolation layer and a convolution layer, and fuses the intermediate features from the corresponding layer of the encoder through a skip connection to integrate multi-scale local and global information, and finally outputs a high-resolution feature map; The high-resolution feature map is processed by the Sigmoid activation function pixel by pixel to generate the future t f The probability distribution P(x,y,t) of the position (x,y) at each time t in the moments, where the last time t p +t f The probability distribution P(x,y,t p +t f ), as the probability distribution prediction of the future target position of the unmanned boat.

3. The unmanned boat trajectory prediction method based on target drive and dynamic constraints according to claim 1 is characterized in that: The sampling and generation of K possible future target point sets G in step S1. specifically includes: Apply the Soft-argmax function to the target position probability distribution, and select the point with the largest probability value as the first target point; Applying a clustering algorithm to divide the remaining candidate point set into K-1 clusters based on the target position probability distribution, and selecting the center point of each cluster as the target point; Combine the first target point with the target point to form a possible future target point set G of the unmanned boat = {G a }∪ {G1,G2,…,G K-1 }。 4. The unmanned boat trajectory prediction method based on target drive and dynamic constraints according to claim 1 or 3, characterized in that: The dynamic equation of the unmanned boat in step S2. is as follows: Among them, q(0) is the initial state of the unmanned boat; f θ is the control function, which is determined by the current time t, the current state q(t) and the target state q(T). o(t) is the position at the current time t, is the speed at the current time t; α t (t,q t:t-M ) is dynamic randomness, simulating the small disturbance of the ocean environment to the unmanned boat, and depends on the historical state of the past M time steps.

5. The unmanned boat trajectory prediction method based on target drive and dynamic constraints according to claim 4 is characterized in that: The dynamic equation of the unmanned boat in step S2. includes a target driving force term and a dynamic randomness term, wherein the target driving force term is calculated by the relative relationship between the position o(t) at the current time t and the target position o(T) to calculate the expected speed With acceleration The formula is as follows: In the formula, τ represents the time required to adjust the current speed to the desired speed, which is determined by the neural network Learn to get; The dynamic randomness term α t (t,q t:t-M ), which is used to simulate the impact of ocean environment disturbance on the trajectory. The conditional probability distribution can be expressed as: In the formula, z represents the latent variable, which is used to capture α t The distribution shape of , in the latent space, follows a normal distribution; P(z) is the prior distribution of z.

Images