A bamboo forest unmanned aerial vehicle full-autonomous visual navigation method and system fusing human operation experience

Abstract

The present application belongs to the technical field of unmanned aerial vehicle navigation, and relates to a bamboo forest unmanned aerial vehicle full-autonomous visual navigation method and system fusing human operation experience. The system integrates multiple source perception units to construct a three-dimensional scene and a risk map; the pilot's line of sight distribution and operation habit are collected through eye tracking and flight control data, human experience strategy priori is formed through time sequence feature extraction and imitation learning, and is adaptively optimized under a reinforcement learning framework in combination with multi-target rewards such as coverage, safety, energy efficiency and behavior consistency to obtain a transferable autonomous navigation strategy; obstacle avoidance flight is realized in a manner combining global task route planning and local rolling track control, and automatically switches to a local safety strategy when communication or positioning is abnormal. Compared with the prior art, the present application can stably complete bamboo forest autonomous inspection in a complex environment, and significantly improves operation safety, running efficiency and information acquisition quality.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) navigation technology, and relates to a fully autonomous visual navigation method and system for bamboo forest UAVs that integrates human operating experience. Background Technology

[0002] Currently, drones, as important equipment for precise management and understory inspection of bamboo forests, have been initially applied in tasks such as resource surveys, fire prevention patrols, pest and disease monitoring, and infrastructure inspections. However, due to the complexity of the bamboo forest environment, limited communication, and insufficient algorithm intelligence, existing bamboo forest drone operations still mostly rely on manual remote control or semi-autonomous flight modes, lacking fully autonomous perception and decision-making capabilities for complex understory environments.

[0003] On the one hand, bamboo forest environments are characterized by narrow spaces, high vegetation density, numerous pole-shaped obstacles, and intensely uneven lighting. GNSS signals under the forest canopy are easily blocked. Traditional navigation methods based on monocular / binocular vision or simple inertial navigation are prone to target loss, path deviation, and positioning inaccuracies under conditions of low light, high contrast, repetitive textures, and swaying branches and leaves, severely limiting the stability and safety of UAVs flying at low altitudes near the ground. On the other hand, wireless communication links under the forest canopy are easily affected by terrain occlusion and vegetation attenuation, resulting in limited communication bandwidth, significant fluctuations in link quality, and hindering real-time human-machine interaction, making it difficult to support remote, precise control and dynamic manipulation under complex tasks. Furthermore, existing autonomous navigation algorithms generally rely on static rules or single-scene training, lacking the ability to adapt to seasonal changes, bamboo growth variations, and task diversity, and are unable to effectively cope with multi-target, long-term, and highly uncertain inspection tasks in real bamboo forests.

[0004] In terms of human-machine collaboration, although the introduction of technologies such as virtual reality (VR), first-person perspective (FPV), and ground station visualization has improved the immersion and safety of operation to some extent, most existing systems are still at the stage of "human-controlled machines," failing to fully utilize the operational experience, gaze patterns, and decision-making logic of forest rangers or drone pilots to guide the intelligent learning of drones. The visual attention mechanisms, risk assessment experience, and path selection strategies developed by humans in long-term operations in complex forest areas have not yet been effectively digitally expressed and transferred to airborne strategies. As a result, drones still need to rely on a lot of trial-and-error exploration in unknown or complex bamboo forest scenarios, which is inefficient, costly, and poses safety risks.

[0005] To overcome the aforementioned bottlenecks, this invention proposes a fully autonomous visual navigation method for bamboo forest drones that integrates human operational experience. This method integrates the operational experience of forest rangers / pilots with the self-learning capabilities of an AI agent, constructing a knowledge-guided framework based on human-machine co-intelligence. The system utilizes a self-evolutionary mechanism combining multimodal sensor fusion, blind-spot-free and distortion-free gaze acquisition, and reinforcement learning combined with imitation learning. This enables the drone to extract quantifiable experiential features from human gaze trajectories, operational habits, and strategic preferences, completing an intelligent evolution from "experience imitation" to "autonomous reasoning."

[0006] This method enables bamboo forest drones to autonomously perform visual navigation, path planning, and risk avoidance in complex forest environments. Furthermore, it allows for strategy transfer and capability self-evolution during long-term inspection missions, achieving truly fully autonomous bamboo forest inspection operations. This technology breaks through the traditional "human-machine separation" paradigm, constructing a three-element collaborative smart forestry operation system of "forest ranger-AI-environment." It holds significant scientific importance and application value for promoting the autonomous, systematic, and intelligent development of forestry intelligent equipment in my country. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a fully autonomous visual navigation method and system for bamboo forest drones that integrates human operating experience.

[0008] The fully autonomous visual navigation method for bamboo forest drones that integrates human operating experience, provided by this invention, includes the following steps:

[0009] (1) The UAV collects multi-source data, and uses a unified time reference to perform time alignment and labeling on the various collected data to form multimodal observation data:

[0010] During the manual teaching phase, forest rangers or drone pilots control the drone through a head-mounted display or first-person perspective terminal with integrated eye tracking. The system simultaneously records the gaze point, dwell time, and eye movement trajectory, and collects the corresponding flight control commands.

[0011] (2) The spherical reconstruction is performed on the acquired equidistant rectangular projection panoramic image. The spherical field of view is divided into several sub-regions. For each sub-region, a mapping transformation from ERP plane coordinates to spherical coordinates is constructed. The polar stretching error is reduced by optimizing the distortion correction objective function, so that the gaze direction deviation after reconstruction is controlled within a predetermined angle threshold. The corrected gaze points, control vectors and time intervals are organized into a temporal behavior sequence. The behavior sequence is encoded using a long short-term memory network to obtain an empirical embedding vector.

[0012] and the current environment status With empirical embedding vectors Input imitation learning network, learns prior human operational strategies:

[0013]

[0014] in, Indicates the state The strategy output is derived from human experience.

[0015] This refers to choosing an action given a state and empirical characteristics. The probability of;

[0016] The strategy prior Used as an initialization strategy or constraint strategy for subsequent reinforcement learning;

[0017] (3) During the inspection mission, the interaction between the UAV and the environment is modeled as a Markov decision process;

[0018] A two-layer experience storage structure is constructed, comprising an online experience buffer and an expert experience base, and the prior parameters obtained through imitation learning and the parameters updated through reinforcement learning are fused according to weights:

[0019] (4) Construct a 3D environment representation based on the depth / LiDAR and visual mapping results, and discretize it into a set of spatial units;

[0020] A graph structure is constructed between the mission starting point and the target inspection area using spatial units or their center points;

[0021] A global reference flight path is generated using a heuristic graph search algorithm, and rolling optimization or model predictive control is used to correct the local trajectory during the execution phase, so that the actual flight path follows the reference path while satisfying obstacle constraints and control constraints.

[0022] Preferably, after completing a round of inspection tasks, a task-level comprehensive score is constructed based on the task log:

[0023] ;

[0024] in, For coverage metrics, For safety indicators, This is a metric for task efficiency. The information quality index is used as the basis for the iterative update of the strategy parameters, with immediate rewards and task-level comprehensive scores as the joint optimization objectives.

[0025] Preferably, in step (3), during the multi-round task interaction process, the parameters of the policy network are optimized through gradient ascent, and their update can be written as:

[0026] ;

[0027] in, This is the learning rate, with a value ranging from 10. -4 ~10 -2 ; The objective function is the cumulative return. Discount factor; For the gradient operator with respect to the policy parameters, Indicates the first Rotate tasks;

[0028] Instant rewards Constructed as a weighted combination of multiple indicators:

[0029] ;

[0030] in, Rewards are given for coverage and path tracking accuracy. As a reward for safe obstacle avoidance, Rewards for energy efficiency Rewards are based on the similarity between the strategy and human experience. , , , These are the weighting coefficients.

[0031] Preferably, the local trajectory rolling optimization control uses the predicted time domain length. Control time domain length In addition to the tracking error term and the control increment term, the cost function also includes a nonlinear penalty term related to the distance to obstacles.

[0032] Preferably, when the communication quality or GNSS / RTK positioning quality is detected to be continuously below a threshold within a preset time window, the system automatically switches to the local security policy network and relies solely on airborne perception to complete preset actions such as safe return or hovering.

[0033] Preferably, the local security policy network is trained offline by selecting a number of inspection data with the highest task-level comprehensive scores, and periodically updated during online operation based on new high-scoring task data.

[0034] Preferably, the reinforcement learning algorithm in step (3) adopts one or more combinations of proximal policy optimization, soft actor-commentator or deep deterministic policy gradient, the discount factor ranges from 0.95 to 0.99, and the entropy regularization term is used to maintain policy exploration capability.

[0035] This invention also provides a fully autonomous visual navigation system for bamboo forest drones that integrates human operational experience. The system consists of a multi-source environmental perception unit, a human experience modeling unit, a reinforcement learning decision-making unit, an autonomous flight control unit, and a mission evaluation and strategy optimization unit. When the system is running, it executes the methods described above to collaboratively achieve fully autonomous visual navigation in the complex environment of bamboo forests.

[0036] This invention integrates the operational experience of forest rangers / drone pilots with the self-learning capabilities of AI agents, constructing a knowledge-guided framework for human-machine co-intelligence. The system utilizes a self-evolutionary mechanism combining multimodal sensor fusion, blind-spot-free and distortion-free gaze acquisition, and reinforcement learning combined with imitation learning. This enables the drone to extract quantifiable experiential features from human gaze trajectories, operational habits, and strategic preferences, completing an intelligent evolution from "experience imitation" to "autonomous reasoning." Through this method, bamboo forest drones can not only autonomously complete visual navigation, path planning, and risk avoidance in complex forest environments, but also achieve strategy transfer and capability self-evolution in long-term inspection missions, thus realizing truly fully autonomous bamboo forest inspection operations. This technology breaks through the traditional "human-machine separation" operational paradigm, constructing a three-element collaborative intelligent forestry operation system of "forest ranger-AI-environment," which has significant scientific and application value for promoting the autonomous, systematic, and intelligent development of forestry intelligent equipment in my country. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall process of the fully autonomous visual navigation method for bamboo forest UAVs that integrates human operating experience in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the fully autonomous visual navigation system for bamboo forest drones that incorporates human operating experience, as described in an embodiment of the present invention. Detailed Implementation

[0039] To facilitate understanding of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and specific examples. The following examples or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0040] The fully autonomous visual navigation method for bamboo forest drones that integrates human operating experience provided by this invention, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0041] Step 1: Forest Understory Multimodal Environment Sensing and Data Acquisition Steps

[0042] This step is performed by the multi-source environmental sensing unit.

[0043] In this embodiment, the bamboo forest drone is equipped with a multimodal environmental perception system for forest understory scenarios, including but not limited to:

[0044] (1) A 360° panoramic or dual fisheye visual acquisition device for acquiring all-round image information under the forest;

[0045] (2) Three-dimensional lidar or depth camera, used to obtain information on the spatial structure and obstacle distribution of bamboo forest;

[0046] (3) Inertial Measurement Unit (IMU) and GNSS / RTK module, used to acquire aircraft attitude, velocity and position information;

[0047] (4) Environmental sensors such as light intensity, wind speed, temperature and humidity are used to record changes in the understory environment.

[0048] The aforementioned multi-source sensor data are aligned and fused using a unified timestamp synchronization mechanism to construct a multimodal observation vector:

[0049]

[0050] in: Indicates time The fusion of observation datasets; This represents a sequence of image frames acquired by a visual acquisition device. Represents a lidar point cloud or depth map; This represents the attitude and motion state quantities output by the inertial navigation and satellite positioning modules; This indicates environmental parameters (including light intensity, wind speed, temperature, and humidity).

[0051] To achieve geometric registration of data from different modalities in a unified coordinate system, the system adopts a three-dimensional coordinate transformation model.

[0052] Let the spatial point in the sensor coordinate system be... The corresponding point in the navigation coordinate system is Its expression is:

[0053] ;

[0054] in: Points in the original coordinate system; A point in the transformed target coordinate system;

[0055] For the yaw angle Pitch angle Roll angle The rotation matrix formed; This is a translation vector used to translate and correct the origin of the coordinate system.

[0056] When it is necessary to project a 3D point cloud onto the camera's imaging plane, a camera intrinsic parameter matrix is ​​further introduced. ,get:

[0057] ;

[0058] in, This represents the set of points projected onto the image coordinate system, used to achieve joint alignment and semantic annotation between the LiDAR point cloud and the visual image.

[0059] During the human teaching and experience collection phase, forest rangers or drone pilots operate and monitor the bamboo forest scene using head-mounted displays (HMDs) or first-person view (FPV) devices with integrated eye-tracking capabilities. The system simultaneously records the following human-computer interaction information:

[0060] gaze point coordinates: , indicating the current gaze position in the current field of view;

[0061] Duration of gaze: This is used to characterize the intensity of attention given to that location in the current task;

[0062] Eye movement trajectory: It consists of a continuous sequence of fixations and is used to describe the dynamic evolution of visual attention.

[0063] Voice and gesture commands: (Including thrust, pitch, yaw, roll, etc. channels), as well as voice or button commands, to reflect human control intentions for the drone.

[0064] Therefore, at time The output comprehensive information status is represented as follows:

[0065] ;

[0066] in, Indicates at time The comprehensive information state set includes environmental perception data, flight status information, and human-computer interaction features, providing a unified, spatiotemporally consistent, high-quality data foundation for subsequent human operational experience modeling and strategy learning. This module realizes the synchronous acquisition and fusion of visual, structural, inertial, and human cognitive information under the forest canopy, and is the front-end support unit of the fully autonomous visual navigation method for bamboo forest UAVs that integrates human operational experience in this invention.

[0067] Step 2: Human Experience Feature Extraction and Policy Prior Modeling

[0068] This step is performed by the human experience modeling unit.

[0069] In this embodiment, a human experience modeling module is constructed to characterize the visual attention and control habits of forest rangers or drone pilots. The function of this module is to transform the implicit experiences generated during forest patrols, such as "where to look, how to fly, and when to adjust," into high-dimensional features and policy priors that can be directly called by the learning algorithm.

[0070] First, for the equirectangular projection (ERP) panoramic images output by head-mounted displays (HMDs) or first-person view (FPV) devices, the system performs spherical field-of-view reconstruction and gaze point geometric correction. To mitigate the distortion caused by polar stretching, the spherical field of view is divided according to latitude and longitude rules. A number of sub-blocks, where the default value is... Each region corresponds to a spherical field-of-view sub-band. For each sub-band... Establish a mapping transformation from ERP planar coordinates to spherical coordinates. Geometric consistency correction is achieved by minimizing the global distortion error function. The distortion optimization objective function is defined as follows:

[0071] ;

[0072] in: Represents the pixel coordinates in spherical coordinates; This represents the pixel point corresponding to the ERP plane coordinates; Subband Nonlinear coordinate mapping from ERP plane to spherical coordinates;

[0073] To compensate for the overall distortion loss, the optimal mapping is obtained through iterative optimization, ensuring that the gaze point orientation error is controlled within a preset threshold (e.g., making the gaze point mapping error less than...). ).

[0074] The corrected gaze information is projected back into the true spherical field of view, resulting in a high-precision, distortion-free set of gaze point trajectories.

[0075] ;

[0076] in, Indicates the pilot was in The main view direction encoding at each moment.

[0077] Simultaneously, the system acquires control commands synchronized with the gaze behavior (such as thrust-weight, pitch, yaw, roll, speed adjustment, etc.) from the flight control channel, along with the time intervals between adjacent commands, constituting a multimodal temporal behavior segment. This information is then organized and formed into a temporal behavior sequence.

[0078] ;

[0079] in: For a moment The location of the fixation point;

[0080] For a moment The control vectors (including channels for throttle, attitude, and heading).

[0081] The time interval between consecutive instructions;

[0082] For length is The sequence of human operational behaviors is used to reflect the temporal correlation structure of "see-think-act".

[0083] This is used to reflect the time-related structure of "seeing-thinking-moving".

[0084] To extract compressed representations from behavioral sequences that can be used by algorithms, this embodiment employs a Long Short-Term Memory (LSTM) network. Modeling is performed. The network input is a sequence. The output is a fixed-dimensional empirical embedding vector, expressed as follows:

[0085] ;

[0086] in, It portrays advanced experience characteristics of drone pilots in bamboo forest patrol missions, such as their focus on rhythm, risk sensitivity, and action patterns.

[0087] LSTM behavioral modeling networks consist of two to four layers of cascaded memory units, with each layer containing 128–512 hidden units, employing a bidirectional structure and The activation function ultimately maps the output to empirical embedding vectors of dimensions 32–128 through a fully connected layer. .

[0088] After obtaining the experience embedding, the system introduces a Behavioral Cloning Network (BCN) to learn the mapping relationship between "state-experience-action". The BCN will then learn the current environment state. With empirical vectors Using the combination of inputs, we learn the distribution of actions preferred by humans in that state and construct policy priors:

[0089]

[0090] in: Indicates the state The strategy output is derived from human experience.

[0091] This refers to choosing an action given a state and empirical characteristics. The probability of.

[0092] The network is trained by minimizing the cross-entropy loss between AI-predicted actions and actual human actions.

[0093] After training is complete, the policy prior is... Used as the initial policy or constraint in the reinforcement learning phase, i.e., as This information is injected into the decision-making network to provide pilot-inspired initial behavioral references for the bamboo forest drone. With this module, the system can transform the navigation experience accumulated by forest rangers and pilots in complex forest environments into transferable and optimizable policy representations without increasing the burden of manual rule design. This significantly shortens the learning convergence time of the drone in new bamboo forest scenarios and improves its adaptability to typical forest conditions such as severe shading, uneven lighting, and complex structures.

[0094] Step 3: Reinforcement Learning Decision-Making and Transfer Based on Human Experience Constraints

[0095] This step is performed by the reinforcement learning decision unit.

[0096] In this embodiment, based on the multimodal environmental perception under the forest and the modeling of human operational experience, a reinforcement learning (RL) framework is introduced to enable the bamboo forest UAV to continuously optimize navigation decisions through "interactive test flights - reward feedback - policy update" and achieve policy transfer and stable self-evolution through human experience constraints.

[0097] First, the decision-making process of bamboo forest drones in forest under-inspection tasks is formalized as a Markov decision process. At that moment The drone is in operation. (Composed of the environmental representation output by the perception module and its own flight status), based on the parameters as follows Policy network Sampling action Instant rewards for environmental feedback and transition to the next state. During multi-round task interactions, the parameters of the policy network are optimized through gradient ascent, and their update can be written as:

[0098] ;

[0099] in, The learning rate can be set to a value of 10. -4 ~10 -2 ; The objective function is the cumulative return. Indicating in strategy and its parameters The expected value under the determined trajectory distribution (or state / action distribution); Discount factor; For the gradient operator with respect to the policy parameters, Indicates the first Rotation of tasks.

[0100] To balance inspection coverage, safe flight, energy usage, and consistency with human experience, this invention will provide immediate rewards. The design is a multi-objective weighted combination:

[0101] ;

[0102] in: To cover rewards related to navigation accuracy, used to encourage drones to effectively approach target paths or key areas;

[0103] These are safety constraints used to penalize high-risk behaviors that involve approaching obstacles such as tree trunks and branches;

[0104] This is an energy consumption constraint term used to limit ineffective maneuvers and excessive energy consumption;

[0105] This is an experience-consistent reward used to measure the correlation between robot behavior and human experience. The degree of similarity; This is a weighting coefficient that can be configured according to different job tasks.

[0106] In practical implementation, the four sub-rewards can be defined as follows:

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] in: For the current location of the drone, This refers to the corresponding reference location or target point on the planned path;

[0112] The distance between the drone and the nearest obstacle (bamboo pole, tree trunk, etc.). This is a safety threshold; For scale parameters, Let be the collision penalty constant; This represents the estimated energy consumption at the current moment. The maximum energy consumption allowed for the task; The experience embedding vector is output by the human operational experience modeling module; This represents the internal features of the policy network in the current state. The cosine similarity between embeddings characterizes the similarity between policy behavior and human experience style.

[0113] To improve sample utilization efficiency and achieve cross-task knowledge transfer, this embodiment designs a two-layer experience storage and replay structure:

[0114] 1. Online experience buffer (short-term memory pool). Used to cache quadruples sampled from recent rounds of the task:

[0115] ;

[0116] in, This refers to the short-term memory window length. During training... We randomly sample small batches of samples and perform offline fitting and policy updates to mitigate instability caused by temporal correlation.

[0117] 2. Expert experience database (long-term memory pool). Used to store high-quality human-computer interaction segments and excellent autonomous flight trajectories that have been evaluated and screened.

[0118] ;

[0119] in, A quality threshold is used to filter representative operational behaviors; a sample is only added to the database when the immediate reward or task-level evaluation meets the set criteria. The expert experience database provides a reusable knowledge base for the Hsinchu forest site, different seasons, and various task types.

[0120] In a two-tier experience storage structure, the capacity of the online experience buffer is not less than A sample of states – actions – rewards – next states, with an expert experience base capacity of no less than [number missing]. We used a sample of samples and adopted a priority empirical sampling strategy based on TD error or task-level scoring to improve sample utilization efficiency.

[0121] At the policy parameter level, this embodiment employs a hybrid update method combining "human prior experience + reinforcement learning convergence strategy". Let the network parameters of the prior policy obtained through imitation learning be denoted as... The policy parameters obtained through reinforcement learning iterations are: Then the fused strategy parameters can be expressed as:

[0122] ;

[0123] in, Fusion network parameters; These are empirical weighting coefficients.

[0124] when When the value is large, the strategy tends to follow human control style, which is suitable for early training or scenarios with extremely high safety requirements; when the value is gradually reduced, the exploration and optimization role of reinforcement learning is enhanced, which is conducive to discovering new strategies that exceed human intuition in complex forest environments.

[0125] Through the aforementioned reinforcement learning and experience transfer mechanisms, bamboo forest drones can inherit the mature experience of forest rangers and pilots in continuous inspection missions, reducing inefficient exploration. On the other hand, relying on reward-driven self-learning capabilities, they can continuously correct and improve navigation strategies in multiple mission cycles, ultimately forming a highly robust, fully autonomous visual navigation system for forest understory that combines human experience characteristics with adaptive evolution capabilities.

[0126] In this invention, the reinforcement learning algorithm employs one or more combinations of Proximal Policy Optimization (PPO), Soft Actor-Commentator (SAC), or Deep Deterministic Policy Gradient (DDPG), with a discount factor ranging from 0.95 to 0.99, and an entropy regularization term used to maintain policy exploration capability.

[0127] Step 4: Autonomous Visual Navigation and Path Planning in a Bamboo Forest Scene

[0128] This step is performed by the autonomous flight control unit.

[0129] This embodiment addresses the online decision-making requirements of "a fully autonomous visual navigation method and system for bamboo forest UAVs that integrates human operational experience," and constructs a navigation and path planning module oriented towards the bamboo forest environment. This module adopts a hierarchical control structure of "global mission route planning + local trajectory refinement + safety strategy switching under abnormal conditions," enabling the UAV to efficiently perform pre-defined inspections in complex forest environments while maintaining stable flight under local disturbances and communication anomalies.

[0130] I. Environmental Discretization and Global Route Generation

[0131] Based on the 3D bamboo forest reconstruction results output by the multimodal perception module, this embodiment first obtains an environmental representation including information such as terrain undulation, bamboo pole distribution, and no-fly zones.

[0132] To facilitate global planning, the environment area is divided into regular grids or sets of voxels:

[0133] ;

[0134] in: For the first One spatial unit; The unit represents the coordinates of the point. This represents the total number of spatial units.

[0135] Each unit comes with tags (flyable, obstacles, buffer zones, etc.) and local risk indicators (such as bamboo pole density, terrain slope, etc.).

[0136] 1. Global Path Planning

[0137] Starting from the task With the target point or several key inspection points To constrain the process, construct a directed graph consisting of spatial unit centers or graph nodes. For any edge Define the comprehensive cost function:

[0138] ;

[0139] in, The distance between nodes is the Euclidean distance. Indicates changes in altitude, used to suppress frequent ascents or descents;

[0140] For measuring local risks when crossing the connecting edge (such as bamboo pole density, minimum safety margin from the tree trunk, etc.).

[0141] This is a weighting factor used to make trade-offs between flight distance, flight altitude, and risk.

[0142] In the figure The minimum cost path from the starting point to each target point can be found using a heuristic search algorithm (such as a cost-based extension of Dijkstra's algorithm). The global task route can be represented as a sequence of nodes:

[0143] ;

[0144] in, , The intermediate nodes are used to constrain the flight corridor of drones, so that they avoid areas with concentrated bamboo poles and no-fly zones.

[0145] The spatial cell resolution is set to 0.1–0.3 m, the environmental map update frequency is 1–5 Hz, and the height weighting coefficient in the cost function is... and risk weight coefficient It can be adjusted online based on real-time wind field estimation, obstacle density, and inspection task priority.

[0146] II. Local Trajectory Generation and Scrolling Optimization Control

[0147] The global flight path provides a macroscopic flight corridor, but unpredictable factors such as wind disturbances, dynamic operating vehicles, or personnel exist in the bamboo forest environment, necessitating dynamic trajectory correction at a local scale. This embodiment employs a rolling optimization control strategy based on a discrete-time model to continuously adjust the UAV's attitude and speed.

[0148] Let the state vector be:

[0149] ;

[0150] The control input is:

[0151] ;

[0152] in, This represents the position of the drone in the world coordinate system. For heading angle; Forward velocity in the plane; Angular velocity of heading; Control variables related to climb / descent speed or altitude commands.

[0153] The simplified discrete-time kinematics of the UAV can be expressed as:

[0154] ;

[0155] in, To control the cycle.

[0156] With a prediction time domain length of Under the rolling optimization framework, the interpolation points of the global flight path within the local window are used. As a reference trajectory, a local cost function is defined:

[0157] + + ;

[0158] in, For at any time The predicted first The state at any given moment; It is a positive definite weight matrix; For reference, the desired state on the flight path; A penalty term related to the distance to obstacles is used to significantly increase the cost when the predicted trajectory approaches a bamboo pole or tree trunk.

[0159] The rolling optimization problem can be written as:

[0160] .

[0161] in, The control sequence to be optimized is given. The optimal sequence obtained only executes the first control input. The next step is to re-optimize and achieve real-time response to local disturbances and dynamic obstacles; To control the time domain. , These are the lower and upper limits of the control quantity, respectively.

[0162] III. Security Policy Switching Under Communication and Location Anomalies

[0163] Considering the severe GNSS signal obstruction and the susceptibility of communication links to terrain and vegetation in forest environments, this embodiment incorporates a safety policy switching logic in the navigation control module. When any of the following conditions is continuously met for more than a preset time window, a safety policy switching logic is implemented. At this time, the system enters downgrade mode:

[0164] 1) The ground station communication data packet loss rate exceeds the threshold or the RTT (round-trip time) remains abnormal;

[0165] 2) GNSS / RTK positioning quality consistently falls below the preset level, relying solely on inertial navigation and visual odometry to maintain relative positioning;

[0166] 3) Significant external disturbances in the mission area (e.g., sudden strong winds) cause the global flight path tracking error to continuously exceed the limit.

[0167] In degraded mode, navigation decisions no longer rely on external commands, but are instead made by the local security policy network. To take control:

[0168]

[0169] in, The value function obtained from training for safe return or hovering avoidance missions, state It includes information such as current relative altitude, attitude, distance to nearby obstacles, and remaining battery power. Depending on different deployment requirements, One or a combination of the following strategies can be executed:

[0170] Slowly ascend along the nearest accessible corridor to above the canopy, then return in a straight line;

[0171] It slowly hovers or stops in a fixed point in a localized open area, waiting for human intervention;

[0172] Return to the takeoff point or relay station at a conservative speed within the pre-defined "safe corridor".

[0173] Once communication and positioning conditions return to normal and remain stable for a period of time, the system smoothly switches back to the regular global-local planning control mode and writes the flight data from the degraded mode into the log for subsequent offline retraining and optimization of the policy network.

[0174] Through the autonomous visual navigation and path planning module described in this embodiment, the bamboo forest UAV can achieve full-process autonomous flight control in complex forest environments, from task allocation and global corridor planning to local trajectory optimization and safety handling of abnormal working conditions. It forms a closed-loop collaboration with the aforementioned multimodal perception, human experience modeling and reinforcement learning decision-making units, thereby supporting the overall realization of "a fully autonomous visual navigation method and system for bamboo forest UAVs that integrates human operating experience".

[0175] Step 5: Strategy Evaluation and Iterative Optimization

[0176] This step is performed by the task evaluation and strategy optimization unit.

[0177] In this embodiment, after the bamboo forest UAV completes one or more rounds of forest patrol tasks, the navigation strategy is verified and iteratively updated through a combination of task-level performance evaluation and offline retraining, thereby forming a continuous optimization closed loop of "execution-evaluation-retraining-redeployment", which enables the UAV to gradually improve its autonomous decision-making level and environmental adaptability during long-term operation.

[0178] I. Construction of Task-Level Comprehensive Scoring Signal

[0179] After the drone completes a pre-set inspection route or a full operational cycle, the system extracts information such as flight trajectory, attitude changes, obstacle avoidance records, energy consumption, and image data quality from the task log and displays it to the forest ranger for review in a visual interface. Based on the algorithm calculation results and manual verification, a task-level comprehensive score signal is generated. Its form can be expressed as:

[0180] ;

[0181] in: This is an inspection coverage rate indicator used to measure the consistency between the actual coverage area of ​​the drone and the expected inspection area. For safety indicators, factors such as minimum obstacle distance, whether a collision or sudden maneuver will occur are taken into account. It is a mission efficiency indicator, related to total flight time, path redundancy, and energy usage. As an information quality indicator, it is jointly determined by factors such as image sharpness, exposure rationality, and the completeness of key target capture; To map the above four normalized indicators to A comprehensive evaluation function for the interval.

[0182] In this invention, It can be given in the following weighted form:

[0183] ;

[0184] Among them coverage With security As a primary evaluation factor, it occupies a high weight.

[0185] For example, coverage and security can be further quantified as follows:

[0186] ;

[0187] ;

[0188] in, This represents the effective length of the actual inspection path within the target area. For planning reference track length; For the first The distance between the drone and the nearest obstacle (bamboo pole, tree trunk, etc.) at any given moment;

[0189] The set safe distance threshold; This represents the total number of time steps for this task.

[0190] If always maintain ,but This indicates a high safety margin.

[0191] II. Strategy Update Combining Instant Rewards and Task Scoring

[0192] During the inspection mission, the reinforcement learning decision-making module has been based on instant rewards. Perform online or near-line updates on localized behaviors. After the task is completed, a task-level score is generated. It serves as a global performance feedback to participate in policy re-optimization, used to adjust the bias policy network towards "better overall performance".

[0193] Let the first The strategy parameters at the end of the round of tasks are , No. The mission-level rating obtained in each round of missions is Taking into account both time-series immediate rewards and task-level scores, the following update format can be constructed:

[0194] ;

[0195] in, This is a time discount factor used to emphasize the impact of recent behavior on task performance;

[0196] These are the update steps for immediate reward items and task-level rating items, respectively; It can be estimated from the policy gradient or the Actor-Critic structure; The task-level score is treated as a global reward for the "replay sample" and the trajectory segments involved in the task are updated with weights.

[0197] when When the value is large, the strategy update places greater emphasis on the overall effect of the entire task round, which is conducive to forming "task-level planning thinking"; when When the value is larger, the strategy focuses more on local, immediate feedback, which is beneficial for improving the ability to respond to local disturbances. In practical applications, the two types of items can be set differently according to the task being performed (such as fire prevention patrols, facility inspections, or pest and disease surveys).

[0198] III. Selection of High-Performance Strategy Fragments and Construction of a Strategy Library

[0199] After multiple rounds of inspection tasks are completed, the system ranks the strategy performance based on the scoring results of each round and marks the high-scoring tasks and their key trajectory segments in the experience storage module. Let there be a total of... Round of task records, corresponding task-level ratings are Define the score threshold. Construct a high-quality strategy library:

[0200] ;

[0201] in, For the first The policy network corresponding to the end of each task round. The state-action sequences and their empirical embedding vectors from high-quality tasks are preferentially written into the long-term experience base for subsequent offline retraining and distillation.

[0202] Through the By weighted averaging or strategy distillation of the strategies, a robust comprehensive strategy can be obtained that performs well under various understory conditions.

[0203] ;

[0204] in, This represents the expected value.

[0205] This strategy By using the new bamboo forest class or different seasonal conditions as the default deployment strategy, and then making minor adjustments in conjunction with the online reinforcement learning module, rapid deployment of "experience inheritance + scenario adaptation" can be achieved.

[0206] IV. Strategy Stability and Robustness Formation under Long-Term Operation

[0207] As the number of inspection missions increases, the system continuously accumulates flight data under different bamboo age structures, canopy closures, and seasonal visual conditions, and constantly refines the strategy network through the aforementioned mission-level evaluation and iterative update mechanism. The final navigation strategy has the following characteristics:

[0208] 1. Under various forest lighting and shading conditions, the flight path deviation and safe distance fluctuation remained within the preset range, demonstrating good stability;

[0209] 2. It can automatically adjust the patrol trajectory density and flight rhythm according to the mission objectives (such as focusing on patrolling forest edges, fire lines or pipeline passages), demonstrating its adaptability to mission requirements;

[0210] 3. It can still make reasonable decisions in the face of local environmental disturbances that have never occurred before (such as sudden wind disturbances or local tree falls), and is not prone to oscillation or failure.

[0211] Through the strategy evaluation and iterative optimization module described in this embodiment, the bamboo forest UAV can continuously absorb new mission experience during long-term operation and make targeted corrections and integrations to existing navigation strategies, thereby gradually forming a highly robust fully autonomous visual navigation capability that adapts to different forest types and operational needs. After training on no less than 105 rounds of bamboo forest inspection missions, and evaluated using an external validation dataset, the system's path deviation standard deviation under typical forest conditions is no greater than a set threshold, the collision event rate is significantly lower than the initial strategy, and the inspection time per unit area and energy consumption per unit distance are significantly reduced compared to traditional manual remote control or semi-autonomous flight schemes.

Claims

1. A method for full autonomous visual navigation of unmanned aerial vehicles in bamboo forests by fusing human operational experience, characterized in that, Includes the following steps: (1) The UAV collects multi-source data, and uses a unified time reference to perform time alignment and labeling on the various collected data to form multimodal observation data: During the manual teaching phase, forest rangers or drone pilots control the drone through a head-mounted display or first-person perspective terminal with integrated eye tracking. The system simultaneously records the gaze point, dwell time, and eye movement trajectory, and collects the corresponding flight control commands. (2) The spherical reconstruction is performed on the acquired equidistant rectangular projection panoramic image. The spherical field of view is divided into several sub-regions. For each sub-region, a mapping transformation from ERP plane coordinates to spherical coordinates is constructed. The polar stretching error is reduced by optimizing the distortion correction objective function, so that the gaze direction deviation after reconstruction is controlled within a predetermined angle threshold. The corrected gaze points, control vectors and time intervals are organized into a temporal behavior sequence. The behavior sequence is encoded using a long short-term memory network to obtain an empirical embedding vector. and the current environment status With empirical embedding vectors Input imitation learning network, learns prior human operational strategies: ； in, Indicates the state The strategy output is derived from human experience. This refers to choosing an action given a state and empirical characteristics. The probability of; The strategy prior Used as an initialization strategy or constraint strategy for subsequent reinforcement learning; (3) During the inspection mission, the interaction between the UAV and the environment is modeled as a Markov decision process; a two-layer experience storage structure containing an online experience buffer and an expert experience base is constructed, and the prior parameters obtained through imitation learning and the parameters updated by reinforcement learning are fused according to weights: (4) Construct a three-dimensional environment representation based on the depth / LiDAR and visual mapping results, and discretize it into a set of spatial units; construct a graph structure between the task starting point and the target inspection area using spatial units or their center points; A global reference flight path is generated using a heuristic graph search algorithm, and rolling optimization or model predictive control is used to correct the local trajectory during the execution phase, so that the actual flight path follows the reference path while satisfying obstacle constraints and control constraints.

2. The method according to claim 1, characterized in that, After completing a round of inspection tasks, a task-level comprehensive score is constructed based on the task log: ； in, For coverage metrics, For safety indicators, This is a metric for task efficiency. The information quality indicator is used as the joint optimization objective, with immediate rewards and task-level comprehensive scores as the optimization targets, and the strategy parameters are iteratively updated. To map the above four normalized indicators to A comprehensive evaluation function for the interval.

3. The method according to claim 1, characterized in that, During the multi-round task interaction, the parameters of the policy network are optimized through gradient ascent, and their update is written as: ； in, This is the learning rate, with a value ranging from 10. -4 ~10 -2 ; The objective function is the cumulative return. Indicating in strategy and its parameters The expected value under the determined trajectory distribution or state / action distribution; Discount factor; For the gradient operator with respect to the policy parameters, Indicates the first Rotate tasks; Instant rewards Constructed as a weighted combination of multiple indicators: ； in, Rewards are given for coverage and path tracking accuracy. As a reward for safe obstacle avoidance, Rewards for energy efficiency Rewards are based on the similarity between the strategy and human experience. , , , These are the weighting coefficients.

4. The method according to claim 1, characterized in that, Local trajectory rolling optimization control uses predicted time domain length Control time domain length In addition to the tracking error term and the control increment term, the cost function also includes a nonlinear penalty term related to the distance to obstacles.

5. The method according to claim 1, characterized in that, When the communication quality or GNSS / RTK positioning quality is detected to be continuously below the threshold within a preset time window, the system automatically switches to the local safety policy network and relies solely on airborne perception to complete the preset safe return or hovering behavior.

6. The method according to claim 4, characterized in that, The local security policy network is trained offline by selecting several inspection data points with the highest task-level comprehensive scores, and then periodically updated during online operation based on new high-scoring task data.

7. The method according to claim 1, characterized in that, The reinforcement learning algorithm in step (3) employs one or more combinations of proximal policy optimization, soft actor-commentator, or deep deterministic policy gradient, with the discount factor ranging from 0.95 to 0.99, and the entropy regularization term is used to maintain policy exploration capability.

8. A fully autonomous visual navigation system for bamboo forest drones that integrates human operating experience, characterized by: The system consists of a multi-source environmental perception unit, a human experience modeling unit, a reinforcement learning decision-making unit, an autonomous flight control unit, and a mission evaluation and strategy optimization unit. When the system is running, it executes the method described in any one of claims 1-7 to collaboratively achieve fully autonomous visual navigation in the complex environment of bamboo forest.

Images