Security unmanned vehicle path planning method fusing tactical risk map and game prediction

Abstract

The invention discloses a security and protection unmanned vehicle path planning method fusing a tactical risk map and game prediction, and relates to the technical field of intelligent security and protection, and the method comprises the steps: receiving a task path issued by a remote control security and protection unmanned vehicle, carrying out the semantic perception of an environment, and constructing a static tactical risk map; acquiring real-time observation track sequences of potential opponents, inferring target intentions of the opponents on line through a Bayesian inverse reinforcement learning BIRL model, and generating a high-risk region thermodynamic diagram; generating a speed and angular speed control instruction through a path tracking controller based on the balanced deterrent path; and executing a speed and angular speed control instruction, collecting a real-time observation track sequence of a potential opponent, and performing online fine adjustment of a Bayesian inverse reinforcement learning BIRL model to generate an optimized security unmanned vehicle path. According to the invention, the unmanned vehicle can continuously generate an optimized path with both safety and tactical initiative in a dynamic confrontation environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, and in particular to a path planning method for unmanned security vehicles that integrates tactical risk maps and game theory prediction. Background Technology

[0002] Path planning technology for unmanned vehicles (UAVs) that integrates tactical risk mapping and game theory prediction plays a crucial role in today's intelligent security field. With the widespread application of UAVs in critical areas such as airports, border defense, and energy facilities, path planning must not only meet basic obstacle avoidance and passage requirements but also integrate environmental semantics, task situation, and adversarial behavior prediction to achieve proactive defense and intelligent response. Current mainstream methods typically combine environmental mapping, semantic segmentation, and trajectory prediction technologies to construct risk maps and generate safe paths, supporting the autonomous operation of UAVs in complex scenarios.

[0003] In the field of security unmanned vehicle path planning that integrates tactical risk maps and game theory prediction, traditional path planning methods mainly rely on static or quasi-static environmental risk models, lacking dynamic modeling and response mechanisms for potential adversary behaviors. At the same time, traditional path generation processes are mostly aimed at minimizing risks, failing to incorporate the proactive deterrence intent in security tasks into the optimization objective, making it difficult to achieve forward-looking path decisions in adversarial scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a path planning method for security unmanned vehicles that integrates tactical risk maps and game theory prediction to solve the problem of insufficient foresight in path planning under adversarial scenarios caused by the lack of dynamic modeling of adversary behavior and active deterrence mechanisms in traditional methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a path planning method for security unmanned vehicles that integrates tactical risk maps and game theory prediction, comprising: Receive the task path issued by the remotely controlled security unmanned vehicle, perform semantic perception of the environment, and construct a static tactical risk map; Collect real-time observation trajectory sequences of potential adversaries, infer the adversary's target intent online through a Bayesian inverse reinforcement learning (BIRL) model, and generate heat maps of high-risk areas; By spatiotemporally fusing static tactical risk maps with heat maps of high-risk areas, a fused tactical risk map is generated. Based on the fusion of tactical risk map and the adversary's objective intentions, a Stackelberg game is constructed, and combined with the objective function of the security unmanned vehicle, the equilibrium deterrence path is solved; Based on the balanced deterrence path, speed and angular velocity control commands are generated through the path tracking controller; The system executes speed and angular velocity control commands, collects real-time observation trajectory sequences of potential adversaries, performs online fine-tuning of the Bayesian Inverse Reinforcement Learning (BIRL) model, and generates an optimized path for the security unmanned vehicle.

[0007] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps of receiving the task path issued by the remotely controlled security unmanned vehicle, performing semantic perception of the environment, and constructing a static tactical risk map are as follows: The security unmanned vehicle receives a task path composed of geographic coordinates and performs semantic perception of the surrounding environment to acquire point cloud data; Based on point cloud data and task paths, the exposed area, the shelter area, the area near key facilities and the passable road area are obtained through semantic segmentation algorithms. Risk values ​​were calculated for exposed areas, bunker areas, areas adjacent to critical facilities, and accessible road areas, and a static tactical risk map was constructed through weighted fusion.

[0008] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the specific steps for collecting real-time observation trajectory sequences of potential adversaries are as follows: By monitoring unauthorized moving targets in the surrounding environment in real time, potential adversaries can be identified. Collect real-time observation trajectory sequences of potential adversaries.

[0009] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps of inferring the opponent's target intent online using a Bayesian inverse reinforcement learning (BIRL) model and generating a high-risk area heatmap are as follows: Based on the real-time observed trajectory sequence of potential adversaries, the target intent of the adversaries is inferred online through the Bayesian inverse reinforcement learning (BIRL) model, and the posterior probability of the potential adversaries for each key facility is calculated. By using the A-star path planning algorithm to reverse engineer each critical facility, the most likely intrusion path from the current location of a potential adversary to each critical facility can be obtained. Based on the most likely intrusion path, the current location of the adversary within a preset time period is predicted in a rolling manner. Multiple risk prediction results are generated according to the time step and superimposed into a heat map of high-risk areas.

[0010] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps for generating a fused tactical risk map by spatiotemporally fusing a static tactical risk map with a high-risk area heat map are as follows. Spatial fusion of static tactical risk map and high-risk area heat map is performed to generate fused tactical risk map at the current moment; Based on the multi-frame risk prediction results, the static tactical risk map and the high-risk area heat map of each time step are spatially fused to generate the fused tactical risk map of the corresponding time step. The fusion tactical risk maps corresponding to the time steps are arranged in chronological order to form a spatiotemporal joint cost map in the rolling time domain, and are continuously updated to generate the fusion tactical risk map.

[0011] As a preferred embodiment of the security unmanned vehicle path planning method that integrates tactical risk maps and game prediction as described in this invention, the construction of the Stackelberg game is based on the integrated tactical risk map and the opponent's target intention, setting the security unmanned vehicle as the leader in the Stackelberg game, and setting potential opponents as followers in the Stackelberg game.

[0012] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps for solving the equilibrium deterrence path are as follows: The set of feasible paths for a follower to travel from its current location to each key facility is used as the follower's policy space, and the follower's reward function is constructed based on the posterior probability of the target. Based on the fusion tactical risk map, the environmental risks accumulated by security unmanned vehicles traveling along any candidate path are used as path cost terms, and high-risk areas are identified based on the high-risk area heat map to generate deterrence benefit terms. The path cost term and the deterrence benefit term are combined into the objective function of the security unmanned vehicle; Based on the objective function of the security unmanned vehicle, the policy space of the follower, and the reward function of the follower, the D-Star Lite algorithm is used to search for the path that maximizes the objective function of the security unmanned vehicle on the fused tactical risk map; Based on the path maximizing the objective function of security unmanned vehicles, we solve the equilibrium deterrence path under the Stackelberg game.

[0013] Based on the path maximizing the objective function of security unmanned vehicles, we solve the equilibrium deterrence path under the Stackelberg game.

[0014] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps of generating speed and angular velocity control commands through a path tracking controller based on an equilibrium deterrence path are as follows: Based on the equilibrium deterrence path, the local geometric curvature of the path point sequence is calculated as the curvature of the equilibrium deterrence path. The current mission phase is determined based on the spatial distance relationship between the current location of the security drone on the balanced deterrence path and the heat map of the high-risk area. Based on the curvature of the balanced deterrence path and the current mission phase, combined with the driving speed, speed control commands are obtained. Based on the balanced deterrence path and the current pose information of the security unmanned vehicle, a pure tracking algorithm is used to generate angular velocity control commands.

[0015] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps of executing speed and angular velocity control commands, collecting real-time observation trajectory sequences of potential opponents, and performing online fine-tuning of a Bayesian inverse reinforcement learning (BIRL) model are as follows. The security unmanned vehicle is driven along a balanced deterrence path by speed and angular velocity control commands, and continuously monitors unauthorized moving targets in the surrounding environment to obtain real-time observation trajectory sequences of potential adversaries. Based on the real-time observed trajectory sequence of potential adversaries, an online fine-tuning process is performed using a Bayesian inverse reinforcement learning (BIRL) model to re-infer the adversary's target intent and generate an updated heat map of high-risk areas.

[0016] As a preferred embodiment of the security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction described in this invention, the steps for generating the optimized security unmanned vehicle path are as follows: The updated heat map of high-risk areas is spatiotemporally fused with the static tactical risk map to generate an updated fused tactical risk map. Based on the updated fusion tactical risk map and the updated adversary's objective, the Stackelberg game is updated and combined with the objective function of the security unmanned vehicle to solve the updated equilibrium deterrence path and generate the optimized security unmanned vehicle path.

[0017] The beneficial effects of this invention are as follows: By constructing a Stackelberg game based on the fusion of tactical risk map and the opponent's objective intent, and combining it with the objective function of the security unmanned vehicle, the equilibrium deterrence path is solved. This realizes the offensive and defensive game modeling with the security unmanned vehicle as the leader and the opponent as the follower. The deterrence intent is embedded in the objective function, so that the path has both risk avoidance and active deterrence capabilities. This enables the unmanned vehicle to continuously generate optimized paths that combine safety and tactical initiative in dynamic confrontation environments, thereby improving the intelligent response level and mission execution efficiency of security tasks. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart for a path planning method for security unmanned vehicles that integrates tactical risk maps and game theory prediction.

[0020] Figure 2 A flowchart for constructing a static tactical risk map.

[0021] Figure 3 This is a flowchart for generating and merging heat maps of high-risk areas.

[0022] Figure 4 The flowchart for solving the equilibrium deterrence path and tracking control. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-4As one embodiment of the present invention, this embodiment provides a path planning method for security unmanned vehicles that integrates tactical risk maps and game theory prediction, including the following steps: S1. Receive the task path issued by the remotely controlled security unmanned vehicle, perform semantic perception of the environment, and construct a static tactical risk map; The security unmanned vehicle receives a task path composed of geographic coordinates and performs semantic perception of the surrounding environment to acquire point cloud data; Specifically, the security unmanned vehicle receives a task path composed of geographic coordinates from a remotely controlled security unmanned vehicle. The security unmanned vehicle activates the multi-line LiDAR and single-line LiDAR deployed at the front of the vehicle to scan the environment in front of the vehicle and acquire point cloud data. It also activates the multi-line LiDAR and single-line LiDAR deployed at the rear of the vehicle to scan the environment behind the vehicle and acquire point cloud data. The point cloud data from the front and rear are then merged to form a complete point cloud data covering the environment around the vehicle.

[0027] It should be noted that the task path consisting of geographic coordinates includes a start point, intermediate points, and an end point.

[0028] Based on point cloud data and task paths, the exposed area, the shelter area, the area near key facilities and the passable road area are obtained through semantic segmentation algorithms. Specifically, point cloud data is spatially aligned with the geographic coordinates of key facilities included in the task path to obtain a 3D environmental point cloud with task semantics. By analyzing the point cloud distribution density and normal vector direction, sparse, unobstructed point cloud clusters with uninterrupted line of sight are identified and marked as exposed areas. By identifying the point cloud features of walls and vegetation and analyzing the spatial distribution relationship of surrounding point clouds, point cloud clusters with occlusion capabilities near walls and vegetation are identified and marked as shelter areas. The Euclidean distance from the 3D environmental point cloud location with task semantics to the geographic coordinates of key facilities in the task path is calculated, and point cloud areas with distances less than the safety distance threshold of key facilities are marked as key facility proximity areas. The remaining 3D environmental point clouds with task semantics are fitted to the ground plane using a random sampling consensus algorithm. Based on the fitted plane, the point cloud elevation difference is calculated, the point cloud density and continuity are analyzed, and the ground slope is estimated to identify flat ground that meets vehicle passage conditions. Point cloud areas that meet vehicle passage conditions are marked as passable road areas.

[0029] It should be noted that the critical facility proximity zone refers to the point cloud area in the 3D environment point cloud where the remaining, unclassified 3D environment point cloud points with task semantics have an Euclidean distance from the geographic coordinates of the critical facility specified in the task path that is less than a preset critical facility safety distance threshold. Critical facilities refer to fixed targets that require priority protection in security tasks, such as substations, entrances / exits, and warehouses. The location of critical facilities is explicitly specified by the geographic coordinates in the task path.

[0030] The critical facility safety distance threshold is defined based on the security task level, the type of critical facility and its physical protection requirements, and the value is usually in the range of 10-30 meters (10 meters is used for low-sensitivity facilities or space-constrained areas, 20 meters is the general default value to balance security and path flexibility, and 30 meters is used for high-value and high-risk facilities).

[0031] Risk values ​​were calculated for exposed areas, bunker areas, areas adjacent to critical facilities, and accessible road areas, and a static tactical risk map was constructed through weighted fusion.

[0032] Specifically, exposure risk is obtained by assessing the maximum visibility of the exposed area in the global environment using a visibility analysis method, representing the likelihood of the exposed area being observed from the outside; shielding benefit is obtained by processing the standardized distance from each point in the shield area to the nearest shield core using a shield distance inverse ratio function, representing the shielding protection capability of the shield area; mission sensitivity is obtained by normalizing the Euclidean distance to critical facilities, representing the proximity of the critical facility's vicinity to key security targets; the basic passage cost of the passable road area, generated from the lidar obstacle detection results, represents the ease of passage in the passable road area; and the exposure risk, shielding benefit, mission sensitivity, and basic passage cost are weighted and fused to generate a static tactical risk map.

[0033] It should be noted that the inverse ratio function of cover distance refers to the fact that the closer a point in space is to the nearest cover (such as a wall or large obstacle), the greater the cover benefit, that is, the lower the risk of being detected or attacked; conversely, the farther away, the smaller the cover benefit. The value is usually in the range of [0-1].

[0034] S2. Collect real-time observation trajectory sequences of potential adversaries, infer the adversary's target intent online through a Bayesian inverse reinforcement learning (BIRL) model, and generate heat maps of high-risk areas. Real-time monitoring of unauthorized moving targets in the surrounding environment as potential adversaries, and collection of real-time observation trajectory sequences of potential adversaries; Specifically, based on complete point cloud data covering the environment around the vehicle, dynamic point cloud clusters in the complete point cloud data are clustered and their motion state estimated through moving target detection and tracking algorithms. Unauthorized moving targets in the surrounding environment are identified and marked as potential adversaries. Potential adversaries are continuously tracked. The position coordinates of potential adversaries in continuous time steps are recorded through a time synchronization mechanism, and real-time observation trajectory sequences of potential adversaries are collected.

[0035] It should be noted that unauthorized moving targets refer to mobile entities that appear within the security area without being legally identified through remote control or security authentication mechanisms (such as RFID, facial recognition, or authorized access credentials), including but not limited to unauthorized personnel, non-mission vehicles, or other dynamic obstacles that have entered the patrol area without permission.

[0036] The moving target detection algorithm processes complete point cloud data covering the vehicle's surrounding environment generated by multi-line and single-line LiDAR. By analyzing dynamic changes between point cloud frames, it identifies unauthorized moving targets that may represent potential threats. The algorithm outputs a list of unauthorized moving targets. Building upon the moving target detection algorithm, the moving target tracking algorithm continuously associates the identified moving targets with their identities and predicts their states. This ensures that the same target is stably tracked across consecutive time steps, preventing identity changes or loss of tracking. The moving target tracking algorithm generates a real-time observation trajectory sequence for potential adversaries.

[0037] Based on the real-time observed trajectory sequence of potential adversaries, the target intent of the adversaries is inferred online through the Bayesian inverse reinforcement learning (BIRL) model, and the posterior probability of the potential adversaries for each key facility is calculated. Specifically, the state space originates from a general method of discretizing continuous physical environments, inheriting and utilizing static tactical risk maps and similar environmental representations, serving as a grid map of the security area; the action space is based on reasonable assumptions and simplifications regarding the movement capabilities of potential adversaries, representing the feasible movement directions of potential adversaries between adjacent grids; the reward function structure is planned based on security domain knowledge, adopting a linear form, and is the inner product of a feature vector (containing tactical features such as distance to each key facility, exposure level, and concealment availability) and a weight vector. The training data is historical intrusion event data, which includes complete trajectory sequences of past potential adversaries and corresponding real target key facilities, controlled remotely. The mission history database provides the data; using historical intrusion event data, the posterior distribution of the weight vector in the adversary's implicit reward function is learned through the Markov chain Monte Carlo method to obtain a trained Bayesian Inverse Reinforcement Learning (BIRL) model; the real-time observed trajectory sequence of potential adversaries is input into the BIRL model, using the set of key facilities contained in the mission path as the candidate target space, assuming that each key facility is a real target of the potential adversary, the trajectory likelihood of generating the current real-time observed trajectory sequence under each key facility assumption is calculated; based on the adversary's target posterior probability for each key facility, the trajectory likelihood of generating the current real-time observed trajectory sequence under the assumption that the key facility is a real target of the potential adversary is calculated, expressed as: in, This indicates that, assuming the potential adversary's true objective is the first... Key facilities Under these conditions, generate a real-time observation trajectory sequence of currently observed potential adversaries. The trajectory likelihood; Represents a real-time observed trajectory sequence of a potential adversary; Indicates the first A key facility refers to a fixed target (such as a substation or entrance / exit) that needs to be protected in a security mission, and is an element in the set of candidate targets; This represents the normalization constant, used to ensure that the sum of probabilities is 1. Represents trajectory sequence The Middle The state (position coordinates) at each time step; Indicates that the target is Under the assumptions, state The reward value at that location (calculated by the Bayesian Inverse Reinforcement Learning (BIRL) model based on the inner product of tactical features (such as distance to key facilities, exposure level, and occlusion availability) and the weight vector). This represents the total number of time steps in the trajectory sequence.

[0038] Based on trajectory likelihood and a pre-defined uniform prior distribution, the posterior probability of a potential adversary for each key facility is calculated using Bayes' theorem, expressed as: in, This represents the real-time observation trajectory sequence of a potential adversary. Under these conditions, potential adversaries will take the first Key facilities The probability of being the real target; This represents the trajectory likelihood, assuming the potential adversary's true objective is a critical facility. At that time, generate the currently observed trajectory sequence. The possibility; This indicates that before observing any trajectory, the potential adversary chose the first... Key facilities As the initial probability of the target; Indicates the total number of critical facilities; Indicates the observed trajectory The total probability is used to normalize the posterior probability.

[0039] It should be noted that pre-setting refers to configuring an initial target prior distribution for the Bayesian Inverse Reinforcement Learning (BIRL) model by the remote control terminal during the initialization phase, based on the security task requirements.

[0040] By using the A-star path planning algorithm to reverse engineer each critical facility, the most likely intrusion path from the current location of a potential adversary to each critical facility can be obtained. Specifically, the geographical coordinates of each critical facility are used as the starting point of the A-Star path planning algorithm, and the current location of the potential adversary is used as the ending point. The A-Star path planning algorithm is run on the grid map corresponding to the fused tactical risk map. The heuristic function of the A-Star path planning algorithm adopts Euclidean distance, and the cost function is the sum of the fused tactical risk map values ​​of each grid on the path. By maintaining open and closed lists, the node with the smallest total cost (the sum of the fused tactical risk map values ​​of each grid on the path) is continuously selected for expansion until the current location of the potential adversary is reached. The path from the currently processed critical facility to the current location of the potential adversary is obtained by backtracking the parent node chain. The path from the currently processed critical facility to the current location of the potential adversary is reversed to obtain the most probable intrusion path from the current location of the potential adversary to the currently processed critical facility. The above A-Star path planning and path reversal process is performed on all critical facilities one by one to obtain the most probable intrusion path from the current location of the potential adversary to each critical facility.

[0041] It should be noted that the A* path planning algorithm is a classic heuristic search algorithm used to find the optimal path from the starting point to the target point in a known map.

[0042] After the A* algorithm completes the forward search from the critical facility (starting point) to the potential adversary's current location (end point), the path direction obtained is from the critical facility to the adversary. In order to obtain the actual possible invasion direction taken by the adversary (i.e. from the adversary's current location to the critical facility), the node order of the path at the current location needs to be reversed, so that the starting point of the path becomes the adversary's current location and the end point becomes the critical facility, thereby obtaining the most likely invasion path that conforms to the adversary's behavioral logic.

[0043] Based on the most likely intrusion path, the current location of the adversary within a preset time period is predicted in a rolling manner. Multiple risk prediction results are generated according to the time step and superimposed into a heat map of high-risk areas.

[0044] Specifically, the most likely intrusion path is represented as a series of ordered path points. Assuming the potential adversary moves along the most likely intrusion path at a constant speed, the path point sequence is time-parameterized according to a preset time step, mapping the path points to each time step within a preset future time period, and obtaining the predicted location of the potential adversary corresponding to each time step. For the predicted location at each time step, a spatial risk distribution is constructed on the grid map with the predicted location as the center. The spatial risk distribution is smoothed using a Gaussian kernel function to form a single-frame risk prediction result for each time step. Then, the single-frame risk prediction results for all time steps are accumulated grid by grid in the same grid coordinate system to obtain the accumulated risk value. The accumulated risk value is normalized so that all grid values ​​fall within the [0, 1] interval, generating a heat map of high-risk areas.

[0045] It should be noted that the preset time step refers to the fixed time interval (usually 0.5 seconds to 2.0 seconds) used to divide a preset future time period (e.g., 5 seconds) into several discrete time points when making rolling predictions of the future behavior of potential adversaries; the preset future time period refers to a fixed time length (e.g., 5 seconds) pre-configured by the remote control terminal based on the response timeliness of the security scenario, the typical movement speed of the adversary, and the planning cycle task parameters of the unmanned vehicle.

[0046] S3. Spatiotemporally fuse the static tactical risk map with the heat map of high-risk areas to generate a fused tactical risk map; Spatial fusion of static tactical risk map and high-risk area heat map is performed to generate fused tactical risk map at the current moment; Specifically, the static tactical risk map and the current high-risk area heat map are aligned to the same grid map coordinate system to ensure a one-to-one spatial correspondence between the static tactical risk map and the current high-risk area heat map. Each grid location in the grid map is traversed, and the risk value of each grid location in the static tactical risk map and the risk value in the high-risk area heat map are read. The risk values ​​in the static tactical risk map and the risk values ​​in the high-risk area heat map are spatially fused using a weighted superposition formula. Weighted calculations are performed on each grid to obtain the fused risk value of each grid in the fused tactical risk map. After completing the calculations for all grids, a fused tactical risk map with the same resolution as the original map is formed for the current moment.

[0047] Based on the multi-frame risk prediction results, the static tactical risk map and the high-risk area heat map of each time step are spatially fused to generate the fused tactical risk map of the corresponding time step. Specifically, a high-risk area heatmap for the first time step is obtained from the multi-frame risk prediction results, and the high-risk area heatmap is aligned with the static tactical risk map to the same grid map coordinate system. Each grid position in the grid map is traversed, and the risk value of each grid position in the static tactical risk map and the risk value in the high-risk area heatmap for the first time step are read. The risk values ​​in the static tactical risk map and the risk values ​​in the high-risk area heatmap for the first time step are spatially fused using a weighted overlay method to complete the fusion calculation for the first time step and generate the fused tactical risk map for the first time step. The above alignment, traversal, reading, and weighted calculation process is repeated to generate the fused tactical risk map for the second time step. This process is repeated for each time step in the multi-frame risk prediction results, performing the same spatial fusion operation sequentially to finally generate the fused tactical risk map for each time step.

[0048] The fusion tactical risk maps corresponding to the time steps are arranged in chronological order to form a spatiotemporal joint cost map in the rolling time domain, and are continuously updated to generate the fusion tactical risk map.

[0049] Specifically, the fused tactical risk map of the first time step, the fused tactical risk map of the second time step, and so on up to the fused tactical risk map of the last time step are arranged sequentially in chronological order to form a three-dimensional raster data structure containing multiple time layers. This three-dimensional raster data structure with multiple time layers is a rolling time-domain spatiotemporal joint cost map. At each preset time interval, the fused tactical risk map corresponding to the oldest time step in the rolling time-domain spatiotemporal joint cost map is deleted, and the newly generated fused tactical risk map of the next time step is appended to the end of the three-dimensional raster data structure with multiple time layers, completing the rolling update. Through the above arrangement and rolling update operations, the rolling time-domain spatiotemporal joint cost map always covers a prediction window of a fixed length in the future, and the fused tactical risk map corresponding to the current moment is extracted from it as the input for path planning, finally generating the fused tactical risk map.

[0050] It should be noted that the preset time refers to the time that the remote control terminal sets in advance during the task initialization phase based on security response requirements, computing resources, and the speed of typical adversary behavior.

[0051] S4. Based on the fusion of tactical risk map and the opponent's objective intentions, construct a Stackelberg game and combine it with the objective function of the security unmanned vehicle to solve for the equilibrium deterrence path; Based on the fusion of tactical risk map and the target intentions of the adversary, the security unmanned vehicle is set as the leader in the Stackelberg game, and the potential adversaries are set as the followers in the Stackelberg game. Specifically, the Stackelberg game data structure is initialized in the vehicle control computer; the role attribute field of the security unmanned vehicle is set as the leader in the Stackelberg game, and the decision variables of the security unmanned vehicle are bound to the candidate path set of the security unmanned vehicle; the role attribute field of the potential adversary is set as the follower in the Stackelberg game, and the decision variables of the potential adversary are bound to the set of feasible paths from the current location to each key facility; the fused tactical risk map is used as the basis for environmental cost to calculate the cumulative risk of the leader's path, and the posterior probability of the target contained in the adversary's target intention is used as the weight basis for constructing the follower's reward function; based on the Stackelberg game data structure, a sequential decision logic of the leader acting first and the follower responding later is established. The security unmanned vehicle first selects a path, and the potential adversary selects the optimal response path based on its own reward function after observing the path; finally, the roles of the leader and follower in the Stackelberg game are set.

[0052] The set of feasible paths for the follower to reach each key facility from the current location is used as the follower's policy space, and the follower's reward function is constructed based on the posterior probability of the target. Specifically, taking the potential adversary's current location as the starting point and the geographical coordinates of each key facility as the ending point, the A-Star path planning algorithm is used to calculate all feasible paths from the current location to each key facility on the fused tactical risk map. The set of all feasible paths is defined as the follower's policy space. The posterior probability of the potential adversary's target for each key facility is obtained, and the posterior probability of each key facility is output by the Bayesian Inverse Reinforcement Learning (BIRL) model. For each feasible path in the follower's policy space, the key facility pointed to by each feasible path is determined, and the posterior probability of the target for each key facility is used as the reward value of each feasible path. Each feasible path is associated with the reward value of each corresponding feasible path to construct the follower's reward function. The follower's reward function is used to quantify the benefits gained by the potential adversary when choosing a certain feasible path.

[0053] Based on the fusion tactical risk map, the environmental risks accumulated by security unmanned vehicles traveling along any candidate path are used as path cost terms, and high-risk areas are identified based on the high-risk area heat map to generate deterrence benefit terms. Specifically, any candidate path of the security unmanned vehicle is represented as a series of ordered path points. Each path point in this ordered sequence is traversed, and its corresponding risk value in the fused tactical risk map is read. The risk values ​​of all path points are accumulated to obtain the environmental risk accumulated by the security unmanned vehicle traveling along any candidate path, which is used as the path cost term. Based on the high-risk area heatmap, areas with risk values ​​greater than a preset high-risk threshold are identified and defined as high-risk areas. The Euclidean distance from each path point on any candidate path to the nearest high-risk area is calculated. If the Euclidean distance is within a preset safety deterrence range, a deterrence indicator function is activated at any path point, setting its value to 1; otherwise, it is set to 0. The deterrence indicator function values ​​of all path points are accumulated and multiplied by the deterrence weight coefficient to generate a deterrence benefit term.

[0054] It should be noted that the high-risk determination threshold is preset by the remote control terminal during the mission initialization phase. It is used to distinguish between general risk areas and high-risk areas that need to be focused on deterring. The typical value range is 0.5-0.8 (risk value normalized to the [0, 1] interval).

[0055] The path cost term and the deterrence benefit term are combined into the objective function of the security unmanned vehicle; Specifically, the environmental risks accumulated by the security unmanned vehicle traveling along any candidate path are obtained, and the environmental risks are used as the path cost term; the deterrence benefits generated by the candidate path being within the high-risk area's security deterrence range are obtained; the path cost term is negatively converted into a benefit term, and the sum of the negative path cost term and the deterrence benefit term is defined as the security unmanned vehicle objective function. The security unmanned vehicle objective function is used to measure the comprehensive benefits of the security unmanned vehicle when selecting candidate paths, thus completing the construction of the security unmanned vehicle objective function.

[0056] Based on the objective function of the security unmanned vehicle, the policy space of the follower, and the reward function of the follower, the D-Star Lite algorithm is used to search for the path that maximizes the objective function of the security unmanned vehicle on the fused tactical risk map; Specifically, the fused tactical risk map is converted into a grid cost map required by the D-Star Lite algorithm, where the cost of each grid is determined by the risk value of the corresponding location in the fused tactical risk map. The open and closed lists of the D-Star Lite algorithm are initialized, with the starting point of the security drone and the ending point of the balanced deterrence path as the planning start and end points. In each iteration of the D-Star Lite algorithm, the comprehensive benefit of the current candidate path is calculated based on the security drone's objective function; the comprehensive benefit equals the sum of the negative path cost term and the deterrence benefit term. Simultaneously, for each feasible path in the follower's policy space, the optimal response of potential adversaries is evaluated through the follower's reward function, and the adversary's optimal response is fed back into the security drone's objective function to correct the path benefit. The D-Star Lite algorithm dynamically adjusts node priorities based on the corrected benefits, prioritizing the expansion of path directions that maximize the security drone's objective function. This iteration continues until the optimal path from the starting point to the ending point is found; the optimal path is the path that maximizes the security drone's objective function.

[0057] Based on the path maximizing the objective function of security unmanned vehicles, we solve the equilibrium deterrence path under the Stackelberg game.

[0058] Specifically, the path maximizing the objective function of the security drone is used as the leader's policy input in the Stackelberg game. Based on this path, the optimal response path for potential opponents (as followers) after observing this path is calculated. The optimal response path is obtained by selecting the feasible path that maximizes the follower's reward function from the follower's policy space. The leader's path and the follower's optimal response path are combined to form a policy pair. The policy pair is then verified to satisfy the Stackelberg equilibrium condition: the objective function value of the security drone under the policy pair is not less than the objective function value under any other path, and the potential opponent's response path is indeed the optimal choice for the potential opponent given the leader's path. If the equilibrium condition is satisfied, the leader's path is determined as the equilibrium deterrence path in the Stackelberg game. If the equilibrium condition is not satisfied, the D-Star Lite algorithm is returned to re-search for a path that maximizes the security drone's objective function until a path satisfying the equilibrium condition is found. Finally, the equilibrium deterrence path is output.

[0059] It should be noted that the criterion for satisfying the equilibrium condition is based on the theoretical definition of Stackelberg games: whether the leader's path maximizes its objective function among all possible paths, and whether the follower's path represents the optimal response. If the above conditions are not met, then the equilibrium condition is not satisfied.

[0060] S5. Based on the balanced deterrence path, speed and angular velocity control commands are generated through the path tracking controller; Based on the equilibrium deterrence path, the local geometric curvature of the path point sequence is calculated as the curvature of the equilibrium deterrence path. Specifically, the equilibrium deterrence path is represented as an ordered sequence of path points. Three consecutive path points are selected from this sequence to form a local triad. For each local triad, the vector connecting the first and last points and the perpendicular distance from the middle point to the line connecting the first and last points are calculated. Combining this with the arc length between the three points, the local geometric curvature at the path point in the middle position of the local triad is calculated using the curvature approximation formula. The process of selecting three points and calculating curvature is repeated for all the middle points in the path point sequence except for the first and last points, to obtain the local geometric curvature corresponding to each middle point. The local geometric curvatures of all the middle points are arranged in the path order to generate a curvature sequence corresponding to the equilibrium deterrence path, and this curvature sequence is used as the curvature of the equilibrium deterrence path.

[0061] The approximate formula for curvature is expressed as follows: in, This represents the local geometric curvature, a quantified value indicating the degree of path curvature at the current midpoint. The greater the curvature, the more curved the path. This represents the perpendicular distance from the midpoint to the line connecting the first and last points; This represents the straight-line distance between the first and last points; It represents the arc length of the path between three points, which is the sum of the two distances from the first point to the middle point and then to the last point.

[0062] The current mission phase is determined based on the spatial distance relationship between the current location of the security drone on the balanced deterrence path and the heat map of the high-risk area. Specifically, the onboard control computer obtains the current location coordinates of the security drone, traverses all grids in the high-risk area heatmap with risk values ​​greater than the high-risk determination threshold, and marks these grids as high-risk areas. The Euclidean distance from the current location of the security drone to each high-risk area grid is calculated, and the smallest Euclidean distance is selected as the shortest distance from the security drone to the high-risk area. This shortest distance is compared with the lower and upper limits of a preset safety deterrence interval. If the shortest distance is less than or equal to the upper limit of the preset safety deterrence interval and greater than or equal to the lower limit, the current task stage is determined to be the approaching high-risk area stage; if the shortest distance is greater than the upper limit of the preset safety deterrence interval, the current task stage is determined to be the routine patrol stage. Finally, the determination result of the current task stage is output.

[0063] It should be noted that the preset lower and upper limits of the safety deterrence zone refer to the spatial boundaries used to define whether the security unmanned vehicle is in a spatial range where effective deterrence can be implemented while maintaining a safe distance. The determination of the spatial range boundary is based on a comprehensive consideration of multiple dimensions of knowledge, including security strategy requirements, vehicle braking performance, adversary response time, and sensor positioning accuracy. The preset safety deterrence zone is pre-set by the remote control terminal during the mission initialization phase based on the security strategy, vehicle braking performance, and adversary response time. The lower limit is used to prevent the unmanned vehicle from getting too close to high-risk areas and falling into danger, while the upper limit is used to ensure that the deterrent behavior has visibility and psychological impact on potential adversaries. The preset safety deterrence zone is usually set between 5 meters and 15 meters, with the lower limit not lower than 5 meters and the upper limit not higher than 15 meters. The specific values ​​can be dynamically adjusted according to the sensitivity of the mission.

[0064] Based on the curvature of the balanced deterrence path and the current mission phase, combined with the driving speed, speed control commands are obtained. Specifically, based on the curvature sequence of the equilibrium deterrence path and the determination results of the current mission phase; if the current mission phase is a routine patrol phase, the base driving speed is set according to the curvature of the equilibrium deterrence path, with a higher curvature resulting in a lower driving speed and a lower curvature resulting in a higher driving speed; if the current mission phase is approaching a high-risk area, the driving speed is further reduced on top of the base driving speed to ensure that the security unmanned vehicle has sufficient situational awareness and emergency response capabilities; combining the curvature constraints of the equilibrium deterrence path and the requirements of the mission phase, the final speed control command is generated through a speed control algorithm; the speed control command is used to drive the chassis controller to adjust the forward or backward linear speed of the security unmanned vehicle, thereby achieving safe tracking of the equilibrium deterrence path.

[0065] Based on the balanced deterrence path and the current pose information of the security unmanned vehicle, a pure tracking algorithm is used to generate angular velocity control commands.

[0066] Specifically, the balanced deterrence path is represented as a series of ordered path point coordinate sequences. Based on a preset aiming distance, the path point closest to the current position of the security drone and meeting the aiming distance requirement is found within this ordered sequence and designated as the aiming point. The lateral deviation from the current position of the security drone to the aiming point is calculated, defined as the vertical distance between the aiming point and the current heading of the security drone. The heading deviation between the current heading angle of the security drone and the direction of the aiming point is calculated. The lateral deviation and the aiming distance are used to calculate the steering angle using a pure tracking algorithm. Based on the front wheel steering angle and the current speed of the security drone, the steering angle is converted into an angular velocity control command through vehicle kinematics, and an angular velocity control command is generated.

[0067] The pure tracking algorithm expression is: in, This indicates the front wheel steering angle, which refers to the angle at which the front wheels of the unmanned security vehicle need to turn. The wheelbase refers to the distance between the center of the front wheel and the center of the rear wheel of the unmanned security vehicle. The lateral deviation refers to the vertical distance from the aiming point to the longitudinal axis of the vehicle as defined by the current pose of the unmanned security vehicle. The aiming distance refers to the straight-line distance from the current position of the unmanned security vehicle to the aiming point. This represents the arctangent function, used to convert geometric relationships into steering angles, ensuring that the steering angle is within a reasonable range. Typically, its value range is -π / 2 to π / 2.

[0068] It should be noted that the current pose information of the security unmanned vehicle refers to the combination of the vehicle's position coordinates and heading angle at the current moment.

[0069] The pure tracking algorithm selects a preview point on the desired path and moves the autonomous vehicle toward the preview point in its current pose, thereby gradually approaching the entire path.

[0070] Vehicle kinematics refers to the steering angle of the front wheels. With linear velocity Convert to angular velocity control command The mathematical mapping is expressed as: in This indicates an angular velocity control command; This indicates the current speed of the security unmanned vehicle; Indicates the front wheel steering angle; Indicates wheelbase.

[0071] S6 executes speed and angular velocity control commands, collects real-time observation trajectory sequences of potential adversaries, performs online fine-tuning of the Bayesian Inverse Reinforcement Learning (BIRL) model, and generates an optimized security unmanned vehicle path.

[0072] The security unmanned vehicle is driven along a balanced deterrence path by speed and angular velocity control commands, and continuously monitors unauthorized moving targets in the surrounding environment to obtain real-time observation trajectory sequences of potential adversaries. Specifically, the chassis controller receives speed and angular velocity control commands, driving the chassis actuators of the security unmanned vehicle to adjust wheel speed and steering angle, enabling the vehicle to travel along a balanced deterrence path. During travel, the multi-line and single-line lidar deployed at the front of the vehicle are activated to scan the environment in front of and behind the vehicle, acquiring and merging point cloud data to form complete point cloud data covering the surrounding environment. Then, a moving target detection and tracking algorithm is used to identify unauthorized moving targets in the surrounding environment and mark them as potential adversaries. The position coordinates of potential adversaries within continuous time steps are recorded to form a real-time observation trajectory sequence of potential adversaries.

[0073] Based on the real-time observation trajectory sequence of potential adversaries, the model is fine-tuned online using a Bayesian inverse reinforcement learning (BIRL) model, and the adversary's target intent is re-inferred to generate an updated heat map of high-risk areas. Specifically, the real-time observed trajectory sequence of potential adversaries is input into the Bayesian Inverse Reinforcement Learning (BIRL) model. The BIRL model uses the newly added real-time observed trajectory sequence to perform a Bayesian update on the existing posterior distribution of the adversary's implicit reward function, obtaining a fine-tuned posterior distribution of the adversary's implicit reward function. Using the set of key facilities contained in the mission path as the candidate target space, each key facility is assumed to be a real target of the potential adversary. The trajectory likelihood of generating the current real-time observed trajectory sequence under the assumption that the key facility is a real target is calculated. Combined with the uniform prior distribution, the posterior probability of the potential adversary for each key facility is recalculated using the Bayesian formula. Taking each key facility as the starting point and the current position of the potential adversary as the ending point, the most likely invasion path from the current position of the potential adversary to each key facility is reversed using the A-satellite path planning algorithm. The adversary's position within a preset time period is predicted in a rolling manner, generating multi-frame risk prediction results by time step. The multi-frame risk prediction results are spatially superimposed to generate an updated high-risk area heat map.

[0074] The updated heat map of high-risk areas is spatiotemporally fused with the static tactical risk map to generate an updated fused tactical risk map. Specifically, the updated high-risk area heatmap and the static tactical risk map are aligned to the same grid map coordinate system; each grid location in the grid map is traversed, and the risk value of each grid location in the static tactical risk map and the risk value in the updated high-risk area heatmap are read; the risk values ​​in the static tactical risk map and the risk values ​​in the updated high-risk area heatmap are spatiotemporally fused to obtain the fused risk value of each grid in the updated fused tactical risk map; after completing the spatiotemporal fusion of all grids, an updated fused tactical risk map with the same resolution as the original map is generated.

[0075] Based on the updated fusion tactical risk map and the updated adversary's objective, the Stackelberg game is updated and combined with the objective function of the security unmanned vehicle to solve the updated equilibrium deterrence path and generate the optimized security unmanned vehicle path.

[0076] Specifically, based on the leader and followers in a Stackelberg game, the feasible paths of potential adversaries from their current location to each critical facility are used as the follower's policy space. A reward function for the follower is constructed based on the posterior probability of the updated adversary's objective. According to the updated fused tactical risk map, the environmental risk accumulated by the security drone along any candidate path is used as the path cost. High-risk areas are identified based on the updated high-risk area heatmap, and deterrence benefit terms are generated on path segments within a preset security deterrence range from the high-risk areas. The path cost and deterrence benefit terms are combined into the security drone's objective function. The D-Star Lite algorithm is used to search for the path that maximizes the security drone's objective function on the updated fused tactical risk map. The equilibrium deterrence path under the Stackelberg game is solved based on the path that maximizes the security drone's objective function, and the updated equilibrium deterrence path is used as the optimized security drone path.

[0077] In summary, this invention achieves offensive and defensive game modeling with the security unmanned vehicle as the leader and the opponent as the follower by constructing a Stackelberg game based on the fusion of tactical risk map and the opponent's objective intent, and solving for the equilibrium deterrence path by combining it with the objective function of the security unmanned vehicle. By embedding the deterrence intent into the objective function, the path has both risk avoidance and proactive deterrence capabilities, enabling the unmanned vehicle to continuously generate optimized paths that combine safety and tactical initiative in dynamic confrontation environments, thereby improving the intelligent response level and mission execution efficiency of security tasks.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A path planning method for security unmanned vehicles that integrates tactical risk maps and game theory prediction, characterized in that: include, Receive the task path issued by the remotely controlled security unmanned vehicle, perform semantic perception of the environment, and construct a static tactical risk map; Collect real-time observation trajectory sequences of potential adversaries, infer the adversary's target intent online through a Bayesian inverse reinforcement learning (BIRL) model, and generate heat maps of high-risk areas; By spatiotemporally fusing static tactical risk maps with heat maps of high-risk areas, a fused tactical risk map is generated. Based on the fusion of tactical risk map and the adversary's objective intentions, a Stackelberg game is constructed, and combined with the objective function of the security unmanned vehicle, the equilibrium deterrence path is solved; Based on the balanced deterrence path, speed and angular velocity control commands are generated through the path tracking controller; The system executes speed and angular velocity control commands, collects real-time observation trajectory sequences of potential adversaries, performs online fine-tuning of the Bayesian Inverse Reinforcement Learning (BIRL) model, and generates an optimized path for the security unmanned vehicle.

2. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 1, characterized in that: The steps for receiving the task path issued by the remotely controlled security unmanned vehicle, performing semantic perception of the environment, and constructing a static tactical risk map are as follows. The security unmanned vehicle receives a task path composed of geographic coordinates and performs semantic perception of the surrounding environment to acquire point cloud data; Based on point cloud data and task paths, the exposed area, the shelter area, the area near key facilities and the passable road area are obtained through semantic segmentation algorithms. Risk values ​​were calculated for exposed areas, bunker areas, areas adjacent to critical facilities, and accessible road areas, and a static tactical risk map was constructed through weighted fusion.

3. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 2, characterized in that: The specific steps for collecting real-time observation trajectory sequences of potential adversaries are as follows. By monitoring unauthorized moving targets in the surrounding environment in real time, potential adversaries can be identified. Collect real-time observation trajectory sequences of potential adversaries.

4. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 3, characterized in that: The steps for inferring the opponent's target intent online using a Bayesian inverse reinforcement learning (BIRL) model and generating a heatmap of high-risk areas are as follows. Based on the real-time observed trajectory sequence of potential adversaries, the target intent of the adversaries is inferred online through the Bayesian inverse reinforcement learning (BIRL) model, and the posterior probability of the potential adversaries for each key facility is calculated. By using the A-star path planning algorithm to reverse engineer each critical facility, the most likely intrusion path from the current location of a potential adversary to each critical facility can be obtained. Based on the most likely intrusion path, the current location of the adversary within a preset time period is predicted in a rolling manner. Multiple risk prediction results are generated according to the time step and superimposed into a heat map of high-risk areas.

5. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 4, characterized in that: The steps for spatiotemporally fusing the static tactical risk map with the heat map of high-risk areas to generate a fused tactical risk map are as follows. Spatial fusion of static tactical risk map and high-risk area heat map is performed to generate fused tactical risk map at the current moment; Based on the multi-frame risk prediction results, the static tactical risk map and the high-risk area heat map of each time step are spatially fused to generate the fused tactical risk map of the corresponding time step. The fusion tactical risk maps corresponding to the time steps are arranged in chronological order to form a spatiotemporal joint cost map in the rolling time domain, and are continuously updated to generate the fusion tactical risk map.

6. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 5, characterized in that: The construction of the Stackelberg game involves setting the security unmanned vehicle as the leader in the Stackelberg game and setting potential opponents as followers, based on the fusion of tactical risk map and the opponent's objective intention.

7. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 6, characterized in that: The steps for finding the equilibrium deterrence path are as follows. The set of feasible paths for a follower to travel from its current location to each key facility is used as the follower's policy space, and the follower's reward function is constructed based on the posterior probability of the target. Based on the fusion tactical risk map, the environmental risks accumulated by security unmanned vehicles traveling along any candidate path are used as path cost terms, and high-risk areas are identified based on the high-risk area heat map to generate deterrence benefit terms. The path cost term and the deterrence benefit term are combined into the objective function of the security unmanned vehicle; Based on the objective function of the security unmanned vehicle, the policy space of the follower, and the reward function of the follower, the D-Star Lite algorithm is used to search for the path that maximizes the objective function of the security unmanned vehicle on the fused tactical risk map; Based on the path maximizing the objective function of security unmanned vehicles, we solve the equilibrium deterrence path under the Stackelberg game.

8. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 7, characterized in that: The process of generating speed and angular velocity control commands through a path tracking controller based on a balanced deterrence path is as follows: Based on the equilibrium deterrence path, the local geometric curvature of the path point sequence is calculated as the curvature of the equilibrium deterrence path. The current mission phase is determined based on the spatial distance relationship between the current location of the security drone on the balanced deterrence path and the heat map of the high-risk area. Based on the curvature of the balanced deterrence path and the current mission phase, combined with the driving speed, speed control commands are obtained. Based on the balanced deterrence path and the current pose information of the security unmanned vehicle, a pure tracking algorithm is used to generate angular velocity control commands.

9. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 8, characterized in that: The execution of speed and angular velocity control commands, the acquisition of real-time observation trajectory sequences of potential adversaries, and the online fine-tuning of the Bayesian Inverse Reinforcement Learning (BIRL) model are as follows: The security unmanned vehicle is driven along a balanced deterrence path by speed and angular velocity control commands, and continuously monitors unauthorized moving targets in the surrounding environment to obtain real-time observation trajectory sequences of potential adversaries. Based on the real-time observed trajectory sequence of potential adversaries, an online fine-tuning process is performed using a Bayesian inverse reinforcement learning (BIRL) model to re-infer the adversary's target intent and generate an updated heat map of high-risk areas.

10. The security unmanned vehicle path planning method integrating tactical risk maps and game theory prediction as described in claim 9, characterized in that: The steps for generating the optimized security unmanned vehicle path are as follows: The updated heat map of high-risk areas is spatiotemporally fused with the static tactical risk map to generate an updated fused tactical risk map. Based on the updated fusion tactical risk map and the updated adversary objectives, the Stackelberg game is updated, and the updated equilibrium deterrence path is solved to generate an optimized security unmanned vehicle path.

Images