Path planning method and device based on deep learning, and electronic equipment

Abstract

The invention relates to a path planning method and device based on deep learning, electronic equipment and a storage medium, and the method comprises the steps: obtaining an unmanned ship navigation path data set, and generating a target network model based on the unmanned ship navigation path data set; utilizing the target network model to obtain a new node based on probability weighted sampling and path tree extension guided by a feasible region; judging whether the new node meets a constraint condition or not; under the condition that the new node meets the constraint condition, adding the new node into a path tree so as to update the path tree; and performing path optimization and track generation by using the updated path tree. According to the method, a brand-new intelligent sampling method with learning ability can be provided, and the fundamental limitation of path planning efficiency is broken through.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a path planning method, apparatus, electronic device and storage medium based on deep learning. Background Technology

[0002] Unmanned surface vessels (USVs), as intelligent maritime operational platforms, have demonstrated broad and in-depth application potential in several key areas, including marine resource exploration, ecological environment monitoring, and maritime security patrol. The core foundation for their practical application lies in their efficient and reliable autonomous path planning capabilities. Among numerous path planning algorithms, the Rapid Expanding Random Tree (RRT) and its asymptotically optimal improved version stand out. Algorithms, due to their ability to effectively handle complex non-convex spaces containing obstacles and high-dimensional motion constraints, have become one of the mainstream solutions in the field of autonomous navigation for unmanned vessels.

[0003] However, The overall performance of the algorithm is highly dependent on the efficiency of the sampling process. Traditional The algorithm employs a strategy of uniform random sampling across the entire configuration space. However, in environments like the ocean, where most areas are unobstructed open water, the majority of randomly generated sample points are located far from the ideal route. These sample points contribute very little to path construction and are considered invalid or inefficient samples. Consequently, the algorithm needs to accumulate a large number of samples and expand iteratively to generate an initial feasible path, resulting in extremely slow convergence to the optimal solution and a severe waste of computational resources and time.

[0004] To address this issue, existing research has proposed several improvement strategies, such as introducing heuristic rules to bias sample points towards the target point with a certain probability during sampling. However, these methods typically rely on manually preset fixed rules, resulting in simple structures and poor adaptability. They struggle to cope with the complex and ever-changing conditions of real-world marine environments, such as the need to dynamically avoid large islands, reef areas, or waters affected by strong currents, thus limiting their intelligence and adaptability. In complex application scenarios, the efficiency improvement effect of such methods is not significant. Fundamentally, existing technologies have not effectively solved a core problem: how to enable algorithms to possess "intelligent prediction" capabilities, accurately determine the area where the optimal path is most likely to occur, and proactively and adaptively concentrate computational resources.

[0005] Therefore, there is an urgent need to develop a path planning method, device, electronic device, and storage medium based on deep learning to solve one or more of the aforementioned problems. Summary of the Invention

[0006] In view of this, in order to solve the above-mentioned technical problems or some of the technical problems, embodiments of the present invention provide a path planning method, apparatus, electronic device and storage medium based on deep learning.

[0007] Firstly, this application provides a path planning method based on deep learning, including:

[0008] Obtain the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset;

[0009] Using the target network model, new nodes are obtained based on probability-weighted sampling guided by feasible regions and path tree expansion;

[0010] Determine whether the new node meets the constraints;

[0011] If the new node meets the constraints, add the new node to the path tree to update the path tree;

[0012] The updated path tree is used for path optimization and track generation.

[0013] In one possible implementation, generating the target network model based on the unmanned vessel navigation path dataset includes:

[0014] Each path in the unmanned vessel navigation path dataset is processed in a targeted manner to generate a ground truth probability map with the same size as the corresponding environmental map.

[0015] A generative adversarial network containing a generator and a discriminator is constructed using the truth probability map.

[0016] The standard adversarial training process is performed on the constructed adversarial network until the network converges, resulting in a feasible domain map generator with stable performance.

[0017] Generate the target network model using a feasible region map generator.

[0018] In one possible implementation, the step of performing targeted processing on each path in the unmanned vessel navigation path dataset to generate a ground truth probability map with the same size as the corresponding environmental map includes:

[0019] Obtain the mapping relationship between each path in the unmanned vessel navigation path dataset and the environmental map;

[0020] Based on the mapping relationship, mark the areas traversed in each path of the unmanned vessel navigation path dataset;

[0021] For areas that are passed through and areas that are not passed through, preset high probabilities and preset low probabilities are assigned respectively, and a true probability map with the same size as the corresponding environment map is generated.

[0022] In one possible implementation, obtaining new nodes using the target network model based on feasible region-guided probability-weighted sampling and path tree expansion includes:

[0023] The target network model is used to generate a feasible region probability map that matches the current task scenario;

[0024] start up The algorithm's single iteration process involves performing a probability-weighted sampling operation to obtain a set of sampling points;

[0025] Select a target point from the set of sampling points, and find the node closest to the target point in the current path tree, and use it as a new node.

[0026] In one possible implementation, determining whether the new node meets the constraints includes:

[0027] The dynamic collision avoidance decision based on COLREGs constraints determines whether the new node meets the constraint conditions.

[0028] In one possible implementation, the method includes:

[0029] If the new node meets the constraints, the COLREGs decision module is immediately activated so that it generates a local avoidance path that meets collision avoidance compliance and ensures that the path can effectively avoid dynamic obstacle vessels.

[0030] In one possible implementation, the step of using the updated path tree for path optimization and track generation includes:

[0031] Calculate the total path cost of the new node in the updated path tree and its related nodes in the updated path tree;

[0032] Traverse the path tree to neighboring nodes whose distance from the new node meets a preset threshold, and determine whether there are any neighboring nodes that have a lower path cost when reached through the new node;

[0033] If there are adjacent nodes with lower path costs, update the connection relationships between the corresponding nodes to optimize the path tree structure;

[0034] After the reroute optimization is completed, a better path from the origin to the destination is selected, and the discrete waypoint sequence on the optimal path is extracted.

[0035] The discrete waypoint sequence is input into the path smoothing module, which then performs curve fitting on the discrete waypoints and outputs a smooth track with continuous curvature after fitting, ensuring that the track conforms to the ship's dynamic characteristics and actual navigation requirements.

[0036] Secondly, embodiments of this application provide a path planning device based on deep learning, comprising:

[0037] The acquisition module is used to acquire the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset.

[0038] The extension module is used to obtain new nodes by utilizing the target network model and probabilistic weighted sampling and path tree expansion guided by the feasible region;

[0039] The judgment module is used to determine whether the new node meets the constraints.

[0040] An update module is used to add the new node to the path tree when the new node meets the constraints, so as to update the path tree;

[0041] The generation module is used to optimize routes and generate tracks using the updated path tree.

[0042] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the steps of the deep learning-based path planning method described above.

[0043] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the deep learning-based path planning method described above.

[0044] The technical solutions provided in this application have the following advantages compared with the prior art.

[0045] Fundamentally solving the problem of blind search: This invention generates a global feasible region probability map through GAN, which paves the way for subsequent... Search provides powerful "prior knowledge," transforming path planning from aimless exploration into targeted verification. This fundamentally solves the blindness problem of traditional algorithms and improves planning efficiency by orders of magnitude.

[0046] High level of navigation safety and compliance: By deeply integrating the COLREGs decision module, this invention ensures that while pursuing efficiency, the unmanned vessel's collision avoidance behavior strictly complies with internationally accepted navigation "traffic regulations," enabling it to safely and predictably integrate into real shipping traffic flow and greatly reducing the risk of collision.

[0047] Significantly improved path quality and feasibility: Combined with the online path smoothing module, the final output trajectory of this invention is not only geometrically optimal but also dynamically feasible, avoiding sharp turns, ensuring smooth and stable navigation, and reducing unnecessary energy consumption and equipment wear.

[0048] It possesses excellent adaptive and learning capabilities: Since the core sampling guidance module is based on a machine learning model, the system of this invention can continuously collect new navigation data for retraining, thereby achieving continuous performance iteration and self-evolution, and has a stronger ability to adapt to new environments and new tasks. Attached Figure Description

[0049] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0052] Figure 1 A flowchart illustrating a deep learning-based path planning method provided in this application embodiment. Figure 1 ;

[0053] Figure 2 A flowchart illustrating a deep learning-based path planning method provided in this application embodiment. Figure 2 ;

[0054] Figure 3 A flowchart illustrating a deep learning-based path planning method provided in this application embodiment. Figure 3 ;

[0055] Figure 4 A flowchart illustrating a deep learning-based path planning method provided in this application embodiment. Figure 4 ;

[0056] Figure 5 A flowchart illustrating a deep learning-based path planning method provided in this application embodiment. Figure 5

[0057] Figure 6 An example diagram illustrating a path planning method based on deep learning, provided for an embodiment of this application;

[0058] Figure 7 A feasibility probability distribution diagram provided for embodiments of this application;

[0059] Figure 8 A schematic diagram of a path planning system based on deep learning provided in an embodiment of this application;

[0060] Figure 9 A schematic diagram of a path planning device based on deep learning provided in an embodiment of this application;

[0061] Figure 10 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0064] To address the existing technical problem of how to enable algorithms to have "intelligent prediction" capabilities, accurately determine the region where the optimal path is most likely to occur, and actively and adaptively concentrate computing resources, this application proposes a path planning method, device, electronic device, and storage medium based on deep learning.

[0065] Specifically, such as Figure 1 As shown, this application provides a path planning method based on deep learning, including:

[0066] S101: Obtain the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset;

[0067] S102: Using the target network model, new nodes are obtained based on probability-weighted sampling guided by feasible regions and path tree expansion;

[0068] S103: Determine whether the new node meets the constraints;

[0069] S104: If the new node meets the constraints, add the new node to the path tree to update the path tree;

[0070] S105: Utilize the updated path tree for path optimization and track generation.

[0071] In this embodiment, the detailed steps of the method are as follows:

[0072] First, a dataset containing multiple sets of historical navigation paths of unmanned vessels is obtained. Based on this dataset, a target network model is trained and constructed using a deep learning network structure. This model is able to learn the complex mapping relationship between path features and navigation environment.

[0073] Subsequently, using the trained target network model and combining the spatial prior information of the feasible region, a probability-weighted sampling strategy is implemented to guide the expansion process of the path tree in space and generate new path nodes.

[0074] Furthermore, the newly generated nodes are subjected to constraint compliance checks, including checking whether the node is in a safe navigation area, whether it meets kinematic and dynamic constraints, and whether it complies with the navigation rules specified in the mission.

[0075] If a new node satisfies all the preset constraints, it is formally added to the path tree structure, thereby realizing the dynamic updating and expansion of the current path tree;

[0076] Finally, based on the updated path tree structure, a multi-objective optimization algorithm is executed to generate a smooth and executable final flight trajectory, completing the entire generation process from path planning to the actual flight path.

[0077] like Figure 2 As shown, in one possible implementation, generating the target network model based on the unmanned vessel navigation path dataset includes:

[0078] S201: Perform targeted processing on each path in the unmanned vessel navigation path dataset to generate a ground truth probability map with the same size as the corresponding environmental map.

[0079] S202: Construct a generative adversarial network containing a generator and a discriminator using the truth probability map;

[0080] S203: Perform standard adversarial training on the constructed adversarial network until the network converges, and obtain a feasible domain map generator with stable performance.

[0081] S204: Generate the target network model using the feasible region map generator.

[0082] In this embodiment, the process of generating the target network model based on the unmanned vessel navigation path dataset includes the following detailed steps:

[0083] First, each navigation path in the unmanned vessel navigation path dataset is preprocessed in a targeted manner to extract key location points and navigation features in the path. Then, based on the spatial structure and resolution information of the environmental map, a ground truth probability map with a size that perfectly matches the environmental map is generated. In this map, each pixel value represents the probability that the corresponding area is selected by the unmanned vessel.

[0084] Next, using the ground truth probability map as training samples, a generative adversarial network architecture consisting of a generator and a discriminator is constructed. The generator is responsible for simulating and generating a probability map based on the input information, while the discriminator is used to judge the difference between the generated map and the real map.

[0085] Subsequently, the standard adversarial training process was performed on the constructed generative adversarial network. The parameters of the generator and discriminator were optimized through multiple iterations, and the network weights were continuously adjusted until the entire network reached a convergent state. Finally, a feasible domain map generator with stable performance and high accuracy in simulating the real navigation environment was obtained.

[0086] Finally, using the trained feasible domain map generator, combined with actual task requirements and real-time environmental information, a target network model that meets the expected functional and accuracy requirements is generated for subsequent path planning and navigation decisions.

[0087] like Figure 3 As shown, in one possible implementation, the step of performing targeted processing on each path in the unmanned vessel navigation path dataset to generate a ground truth probability map with the same size as the corresponding environmental map includes:

[0088] S301: Obtain the mapping relationship between each path in the unmanned vessel navigation path dataset and the environmental map;

[0089] S302: Based on the mapping relationship, mark the areas traversed in each path of the unmanned vessel navigation path dataset;

[0090] S303: Assign preset high probability and preset low probability to the passed area and the non-passed area respectively, and generate a true probability map with the same size as the corresponding environment map.

[0091] Example: Offline training of Generative Adversarial Networks (GANs) – this step is a preparatory phase before planning. First, a large-scale, high-quality dataset of unmanned surface vessel (USV) navigation paths is collected or generated. For each expert path, processing is performed to generate a corresponding "ground truth map". The map is the same size as the environment map, with areas traversed by the path having high probability values ​​and other areas having low probability values. Then, a generative adversarial network (GAN) is constructed, consisting of two competing deep neural networks:

[0092] Regarding the generator (G), its input is a random noise. acoustic vector (Or, combined with a conditional vector of task information), its task is to learn and output a feasible region probability map of the same size as the environment map. .

[0093] The discriminator (D) takes a probability map as input and its task is to determine whether the map is real. Or is it faked by the generator G? Through standard adversarial training, the two networks engage in a "minimum-maximum" game, the objective function V(D,G) of which can be expressed as:

[0094] ;

[0095] After training converges, a powerful feasible region map generator G is obtained, which has learned the spatial probability distribution of the optimal path in a given task environment.

[0096] like Figure 5 As shown, in one possible implementation, the step of obtaining new nodes using the target network model based on feasible region-guided probability-weighted sampling and path tree expansion includes:

[0097] S401: Generate a feasible domain probability map that matches the current task scenario using the target network model;

[0098] S402: Start The algorithm's single iteration process involves performing a probability-weighted sampling operation to obtain a set of sampling points;

[0099] S403: Select a target point from the set of sampling points and find the node closest to the target point in the current path tree as the new node.

[0100] For example, probability-weighted sampling and path tree expansion based on feasible regions serve as the online planning phase when an unmanned surface vessel (USV) performs a mission.

[0101] Generate a feasible region map: Call the generator G trained in step one, input the current task information (such as start point, end point, and environment summary), and generate a feasible region probability map for the current task. .

[0102] Probability-weighted sampling: in In each iteration of the algorithm, instead of performing globally uniform random sampling, probability-weighted sampling is performed across the entire map space. For any point... The probability of being selected as a sampling point Rather than in the feasible region probability map The corresponding probability value Proportional:

[0103] ;

[0104] In this way, the sampling points will fall within the optimal path region predicted by the GAN with a high probability.

[0105] Extension and Collision Detection: Select a point from the sampled points, denoted as... Find the nearest node in the current path tree. and to Expanding the fixed step size to obtain a new node Check if the extended path collides with static obstacles.

[0106] In one possible implementation, determining whether the new node meets the constraints includes:

[0107] The dynamic collision avoidance decision-making based on COLREGs constraints determines whether the new node meets the constraints. In one possible implementation, the process of determining whether the new node meets the constraints is accomplished through a dynamic collision avoidance decision-making mechanism built upon the International Regulations for Preventing Collisions at Sea (COLREGs). Specifically, this mechanism analyzes the position, speed, and heading of the new node in real time, based on the ship encounter situations, navigational responsibilities, and avoidance action principles stipulated in COLREGs, and combines this with dynamic environmental factors (such as obstacle distribution, hydrological and meteorological conditions), thereby systematically and efficiently assessing whether the node meets the safety and compliance requirements of the current navigation mission. During the assessment, different ship encounter situations are carefully considered. For example, when a head-on encounter occurs, the mechanism strictly follows the requirements of COLREGs to determine whether the new node's route allows both ships to turn to starboard for safe passage. In the event of a cross-road encounter, the relative positions and headings of the two vessels will be used to determine which vessel is the yielding vessel and which is the straight-ahead vessel. The analysis will then determine whether the new intersection complies with the rules that the yielding vessel should take evasive action and the straight-ahead vessel should maintain its heading and speed.

[0108] Environmental dynamics play a crucial role in the assessment process. The distribution of obstacles affects the feasibility of a new node. If a new node is located close to static obstacles, even if it meets collision avoidance requirements in theory, it may increase the risk of collision in practice, thus rendering the new node ineligible. Hydrological and meteorological conditions, such as water flow speed and direction, and wind direction, also influence the actual navigation trajectory of ships. For example, strong currents may cause ships to deviate from their planned course. Therefore, these factors must be taken into account when determining the location of a new node to ensure safe and compliant navigation under actual conditions.

[0109] Furthermore, this dynamic collision avoidance decision-making mechanism also considers the vessel's own performance characteristics. Different types of vessels have different maneuverability; for example, large oil tankers have a larger turning radius and slower acceleration and deceleration, while small speedboats are relatively agile. When determining whether a new node meets the constraints, the specific performance of the vessel is taken into account to assess whether it can reach the new node as planned and execute the corresponding avoidance actions. If the new node requires the vessel to complete a sharp turn or high-speed acceleration in a short period of time, and the vessel's performance cannot meet this requirement, then this new node does not meet the constraints.

[0110] Through such a comprehensive and systematic evaluation process, dynamic collision avoidance decision-making based on COLREGs constraints can accurately determine whether a new node meets the constraint conditions, providing a reliable basis for path planning, thereby ensuring that ships navigate safely and compliantly in complex maritime environments.

[0111] In one possible implementation, the method includes:

[0112] If the new node meets the constraints, the COLREGs decision module is immediately activated so that it generates a local avoidance path that meets collision avoidance compliance and ensures that the path can effectively avoid dynamic obstacle vessels.

[0113] Dynamic collision avoidance decision-making based on COLREGs constraints in the generation of new nodes During the process, dynamic collision avoidance checks are performed in parallel. If a collision risk with a vessel with a dynamic obstacle is predicted, the COLREGs decision module is activated. This module generates a local avoidance path that conforms to maritime collision avoidance rules based on the encounter scenario, and uses the endpoint of this path as the final new node. During the process, dynamic collision avoidance checks are performed in parallel. If a collision risk with a vessel with a dynamic obstacle is predicted, the COLREGs decision module is activated. This module generates a local avoidance path that conforms to maritime collision avoidance rules based on the encounter scenario, and uses the endpoint of this path as the final new node. During the process, dynamic collision avoidance checks are performed in parallel. If a collision risk with a vessel with a dynamic obstacle is predicted, the COLREGs decision module is activated. This module generates a local avoidance path that conforms to maritime collision avoidance rules based on the encounter scenario, and uses the endpoint of this path as the final new node.

[0114] like Figure 5 As shown, in one possible implementation, the step of using the updated path tree for path optimization and track generation includes:

[0115] S501: Calculate the total path cost of the new node in the updated path tree and related nodes in the updated path tree;

[0116] S502: Traverse the path tree to neighboring nodes whose distance from the new node meets a preset threshold, and determine whether there are any neighboring nodes whose path cost is lower when reached through the new node;

[0117] S503: If there are adjacent nodes with lower path costs, update the connection relationships between the corresponding nodes to optimize the path tree structure;

[0118] S504: After the reroute optimization is completed, select the better path from the starting point to the ending point, and extract the discrete waypoint sequence on the optimal path;

[0119] S505: Input the discrete waypoint sequence into the path smoothing module so that the path smoothing module can perform curve fitting on the discrete waypoints and output a smooth track with continuous curvature after fitting, ensuring that the track meets the ship's dynamic characteristics and actual navigation requirements.

[0120] For example, path tree optimization and track generation include the following steps:

[0121] Step 1: Create a new node that meets all constraints. Add to the path tree and execute. Algorithm cost calculation and rewire optimization. Path cost function. Taking into account factors such as path length, energy consumption, and risk:

[0122] ;

[0123] Step 2: Once a better path is found, sequence the discrete waypoints along that path. The data is fed into the path smoothing module. This module uses B-spline curve fitting technology to generate a smooth path with continuous curvature. Its definition is:

[0124] ;

[0125] The generated curve It conforms to the characteristics of ship dynamics and can be directly executed by the ship's underlying control system.

[0126] like Figure 6 As shown, the path planning system 100 proposed in this application includes: a sensor module 101, a data processing and storage module 102, and a core planning processor 103.

[0127] like Figure 7 As shown, the specific implementation process of the method in this application is as follows:

[0128] At the start of a planning task (step S701), the core planning processor 103 activates the GAN feasible region generation module (e.g., Figure 6 (103a in the middle).

[0129] like Figure 8 As shown, this module calls a pre-trained GAN generator G. Generator G generates a feasible region probability map 305 based on the current task (start point 301, end point 302) and the environment map (including obstacles 304). In this map, brighter areas (such as 305a) represent areas with a higher probability of being traversed by the optimal path.

[0130] In the planned iterative loop (step S702), the system activates the probability-weighted sampling module (such as...). Figure 6 (See 103b in the diagram). This module samples based on the probability map 305, and the generated sampling points (such as point cloud 306) are densely distributed in the brighter "feasible region" and rarely appear in the darker areas.

[0131] The system selects one sampling point and searches for the nearest node in the path tree for expansion (steps S703, S704). During the expansion process, COLREGs constraint modules (such as...) Figure 6 103c) is activated in parallel (step S705) to handle the dynamic collision avoidance problem, the principle of which is as follows: Figure 8 As shown.

[0132] Finally, once a path to the destination is found, the sequence of nodes along that path is fed into the path smoothing module (e.g., ...). Figure 6 (Step S707) 103d, output the final executable track (Step S708).

[0133] This application uses GAN to generate feasible region maps to guide [the process]. The sampling method successfully combines the pattern recognition capabilities of artificial intelligence with the rigor of traditional search algorithms, providing a basis for achieving a higher level of autonomous ship navigation.

[0134] like Figure 9 As shown, in a second aspect, embodiments of this application provide a path planning device based on deep learning, comprising:

[0135] The acquisition module 901 is used to acquire the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset.

[0136] Extension module 902 is used to obtain new nodes by using the target network model and probabilistic weighted sampling and path tree expansion guided by feasible region;

[0137] The judgment module 903 is used to determine whether the new node meets the constraint conditions;

[0138] The update module 904 is used to add the new node to the path tree to update the path tree if the new node meets the constraints.

[0139] Module 905 is used for path optimization and track generation using the updated path tree.

[0140] In one possible implementation, the acquisition module 901 is further configured to perform targeted processing on each path in the unmanned vessel navigation path dataset to generate a ground truth probability map with the same size as the corresponding environment map; construct a generative adversarial network (GAN) containing a generator and a discriminator using the ground truth probability map; perform a standard adversarial training process on the constructed adversarial network until the network converges to obtain a feasible region map generator with stable performance; and generate a target network model using the feasible region map generator.

[0141] In one possible implementation, the acquisition module 901 is further configured to acquire the mapping relationship between each path in the unmanned vessel navigation path dataset and the environmental map; mark the areas traversed in each path in the unmanned vessel navigation path dataset according to the mapping relationship; assign preset high probability and preset low probability to the traversed areas and the non-traversed areas respectively, and generate a ground truth probability map with the same size as the corresponding environmental map.

[0142] In one possible implementation, the extension module 902 is further configured to utilize the target network model to generate a feasible domain probability map matching the current task scenario; start The algorithm's single iteration process involves performing a probability-weighted sampling operation to obtain a set of sampling points; selecting a target point from the set of sampling points; and finding the node closest to the target point in the current path tree as the new node.

[0143] In one possible implementation, the judgment module 903 is further used to determine whether the new node meets the constraint conditions in a dynamic collision avoidance decision based on COLREGs constraints.

[0144] In one possible implementation, the generation module 905 is further configured to calculate the total path cost of the new node in the updated path tree and related nodes in the updated path tree; traverse adjacent nodes in the path tree whose distance from the new node meets a preset threshold, and determine whether there are any adjacent nodes with a lower path cost when reached through the new node; if there are adjacent nodes with a lower path cost, update the connection relationship between the corresponding nodes to optimize the path tree structure; after the rewire optimization is completed, select the better path from the start point to the end point, and extract the discrete waypoint sequence on the optimal path; input the discrete waypoint sequence into the path smoothing module, so that the path smoothing module performs curve fitting on the discrete waypoints, and outputs a smooth track with continuous curvature after fitting, ensuring that the track conforms to the ship dynamics characteristics and actual navigation requirements.

[0145] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the steps of the deep learning-based path planning method described above.

[0146] This application embodiment also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the deep learning-based path planning method described above.

[0147] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0148] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A path planning method based on deep learning, characterized in that, include: Obtain the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset; Using the target network model, new nodes are obtained based on probability-weighted sampling guided by feasible regions and path tree expansion; Determine whether the new node meets the constraints; If the new node meets the constraints, add the new node to the path tree to update the path tree; The updated path tree is used for path optimization and track generation.

2. The method according to claim 1, characterized in that, The generation of the target network model based on the unmanned vessel navigation path dataset includes: Each path in the unmanned vessel navigation path dataset is processed in a targeted manner to generate a ground truth probability map with the same size as the corresponding environmental map. A generative adversarial network containing a generator and a discriminator is constructed using the truth probability map. The standard adversarial training process is performed on the constructed adversarial network until the network converges, resulting in a feasible domain map generator with stable performance. Generate the target network model using a feasible region map generator.

3. The method according to claim 2, characterized in that, The step of performing targeted processing on each path in the unmanned vessel navigation path dataset to generate a ground truth probability map with the same size as the corresponding environmental map includes: Obtain the mapping relationship between each path in the unmanned vessel navigation path dataset and the environmental map; Based on the mapping relationship, mark the areas traversed in each path of the unmanned vessel navigation path dataset; For areas that are passed through and areas that are not passed through, preset high probabilities and preset low probabilities are assigned respectively, and a true probability map with the same size as the corresponding environment map is generated.

4. The method according to claim 1, characterized in that, The process of obtaining new nodes using the target network model, based on feasible region-guided probability-weighted sampling and path tree expansion, includes: The target network model is used to generate a feasible region probability map that matches the current task scenario; Initiate a single iteration of the RRT algorithm, perform probability-weighted sampling, and obtain a set of sampling points; Select a target point from the set of sampling points, and find the node closest to the target point in the current path tree, and use it as a new node.

5. The method according to claim 1, characterized in that, The determination of whether the new node meets the constraints includes: The dynamic collision avoidance decision based on COLREGs constraints determines whether the new node meets the constraint conditions.

6. The method according to claim 5, characterized in that, The method includes: If the new node meets the constraints, the COLREGs decision module is immediately activated so that it generates a local avoidance path that meets collision avoidance compliance and ensures that the path can effectively avoid dynamic obstacle vessels.

7. The method according to claim 1, characterized in that, The process of using the updated path tree for path optimization and track generation includes: Calculate the total path cost of the new node in the updated path tree and its related nodes in the updated path tree; Traverse the path tree to neighboring nodes whose distance from the new node meets a preset threshold, and determine whether there are any neighboring nodes that have a lower path cost when reached through the new node; If there are adjacent nodes with lower path costs, update the connection relationships between the corresponding nodes to optimize the path tree structure; After the reroute optimization is completed, a better path from the origin to the destination is selected, and the discrete waypoint sequence on the optimal path is extracted. The discrete waypoint sequence is input into the path smoothing module, which then performs curve fitting on the discrete waypoints and outputs a smooth track with continuous curvature after fitting, ensuring that the track conforms to the ship's dynamic characteristics and actual navigation requirements.

8. A path planning device based on deep learning, characterized in that, include: The acquisition module is used to acquire the unmanned vessel navigation path dataset and generate a target network model based on the unmanned vessel navigation path dataset. The extension module is used to obtain new nodes by utilizing the target network model and probabilistic weighted sampling and path tree expansion guided by the feasible region; The judgment module is used to determine whether the new node meets the constraints. An update module is used to add the new node to the path tree when the new node meets the constraints, so as to update the path tree; The generation module is used to optimize routes and generate tracks using the updated path tree.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based path planning method as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based path planning method as described in any one of claims 1 to 7.

Images