Unmanned aerial vehicle indoor three-dimensional reconstruction autonomous acquisition method and system based on reinforcement learning

Abstract

The invention relates to an unmanned aerial vehicle indoor three-dimensional reconstruction autonomous acquisition method and system based on reinforcement learning, and the method comprises the following steps: constructing a virtual indoor simulation space and a virtual unmanned aerial vehicle model, and laying target acquisition points; a reinforcement learning model is constructed and trained, in the training process, in each time step, the unmanned aerial vehicle obtains target collection point information and inputs the target collection point information into the reinforcement learning model, the reinforcement learning model generates an action instruction and transmits the action instruction to the virtual unmanned aerial vehicle model, and the virtual unmanned aerial vehicle executes geometric transformation, updates the position and orientation and reconstructs view information; and deploying the trained reinforcement learning model in an untrained virtual simulation environment, carrying out an unmanned aerial vehicle autonomous flight experiment and data acquisition, and carrying out quantitative evaluation on the acquisition performance through a plurality of performance indexes. Compared with the prior art, efficient and stable data acquisition can be realized in a complex environment.

Description

Technical Field

[0001] This invention relates to the fields of building information modeling (BIM) and unmanned aerial vehicle (UAV) autonomous control technology, and in particular to a UAV indoor 3D reconstruction autonomous acquisition method and system based on reinforcement learning. Background Technology

[0002] In recent years, the emergence of many image-based intelligent reconstruction algorithms (such as NeRF and 3DGS) has lowered the threshold for data acquisition equipment for reconstruction tasks to sensors such as monocular cameras, greatly reducing the weight of sensor equipment and data transmission requirements. This has promoted the widespread use of lightweight UAVs equipped with sensors as a data acquisition method. These intelligent reconstruction algorithms are characterized by using image data instead of laser point clouds and providing full-field image coverage. However, since each UAV image capture only shows a single position and angle, efficient path planning is necessary to address resource waste. Current research both domestically and internationally has applied algorithms such as PPO and DDPG from deep reinforcement learning to tasks like robot obstacle avoidance and path optimization, mainly focusing on outdoor navigation, target arrival, and path following scenarios. For example, Chinese patent CN115826621A discloses a UAV motion planning method and system based on deep reinforcement learning. However, these methods emphasize "safe arrival" and "obstacle avoidance efficiency," and there are still research gaps in strategy design aspects such as field of view modeling and spatial coverage.

[0003] Against this backdrop, lightweight drones have demonstrated advantages over traditional surveying techniques in confined and complex indoor environments. For example, some studies have found that drones can enter attics through small windows to collect data on surfaces that cannot be covered by ground laser scanning, maintaining 1cm accuracy while reducing equipment costs. Other research indicates that drones can autonomously survey areas without GPS, passing through ventilation windows or collapsed openings, demonstrating greater flexibility and safety. Further studies suggest that drones can save approximately 60% of the time in indoor construction monitoring while maintaining surveying accuracy.

[0004] Therefore, there is currently a lack of a lightweight UAV autonomous data acquisition method with the ultimate goal of three-dimensional spatial reconstruction. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an autonomous acquisition method and system for indoor 3D reconstruction of UAVs based on reinforcement learning, which takes into account both acquisition efficiency and completeness, and provides effective data support for subsequent indoor 3D reconstruction.

[0006] The objective of this invention can be achieved through the following technical solutions: An autonomous data acquisition method for indoor 3D reconstruction using unmanned aerial vehicles (UAVs) based on reinforcement learning, comprising the following steps: S1, Scenario simulation modeling and spatial target point setting: Construct a virtual indoor simulation space and a virtual drone model, and set up target acquisition points; S2, Strategy Learning and Capability Training: Construct and train a reinforcement learning model. The reward function of the reinforcement learning model includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. During training, at each time step, the UAV acquires target collection point information and inputs it into the reinforcement learning model. The reinforcement learning model generates action commands and transmits them to the virtual UAV model. The virtual UAV performs geometric transformations, updates its position and orientation, and reconstructs the field of view information. S3, Autonomous Flight Evaluation and Data Acquisition Implementation: Deploy a trained reinforcement learning model in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and quantitatively evaluate the acquisition performance through multiple performance indicators.

[0007] The construction of the virtual indoor simulation space specifically refers to: The Rhino modeling platform was used to construct a virtual indoor simulation space to simulate typical rooms, corridors and corner nodes in a building. In the modeling process, basic spatial units were first established, and openings, connections and barriers were realized through Boolean operations. The units were then combined in a geometric splicing manner to form the overall environment. Subsequently, furniture and obstacles were introduced, and complex components were simplified by abstracting them into equivalent geometric shapes.

[0008] In the process of modeling the virtual indoor simulation space, the difficulty is adjusted according to three dimensions: number of rooms, functional combination, and boundary obstacles. Specifically, for the number of rooms, the difficulty increases from a single room to a combination of multiple rooms; for functional combination, diverse spaces are formed by connecting corridors and corners; and for boundary obstacles, the complexity of occlusion and obstacle avoidance is increased by adding furniture or obstacles to walls or nodes.

[0009] The method for setting up target acquisition points is as follows: deployment is carried out based on Grasshopper scripts, the exposed surfaces of each component in the simulation space are selected as acquisition targets, and orthogonal point clouds are deployed at preset intervals, wherein the preset intervals are dynamically adjusted according to the complexity.

[0010] In the reinforcement learning model, the environmental state at each time step is represented as a multi-dimensional vector, including the current position of the UAV, the current orientation angle, the number of visible target collection points in the simulated field of view, the coverage rate, and the distance of the UAV to the nearest obstacle in the four directions of front, back, left, and right. The coverage rate is defined as the proportion of the number of visible target collection points in the simulated field of view to the total number of target collection points.

[0011] The actions output by the reinforcement learning model include: Translational movement: Perform a translation of the UAV in any of the forward, backward, left, or right directions with a preset fixed step length; Rotation action: Perform horizontal rotations of the drone by multiples of 45 degrees; Stay put: Remain in place for one cycle, observe or wait for a new strategy; Only one action is allowed to be executed at each time step.

[0012] The reward function of the reinforcement learning model is expressed as: , in, For the first A reward or punishment at any time The instantaneous value, For the first A reward or punishment at any time The weight, n The weights are dynamically adjusted as the number of items to be rewarded or penalized changes during the early and later stages of the training phase. The coverage reward is defined as a preset value multiplied by the percentage of newly visible target collection points within the simulated field of view relative to the total number of target collection points; the new discovery reward is defined as a preset value multiplied by the number of newly discovered target collection points; the exploration reward is defined as a tiered reward, with a first score assigned to the first visit, a second score assigned to the second visit, and no further scores assigned to subsequent visits, wherein the first score is greater than the second score; the out-of-bounds penalty is defined as a preset penalty value assigned when the drone goes out of bounds; the collision penalty is defined as a gradient penalty preset based on the distance between the drone and the obstacle, with a larger penalty value for closer distances; and the efficiency penalty is defined as the number of repeated visits.

[0013] The weights of the coverage reward and exploration reward are positive, with the weight in the early stage being greater than that in the later stage. The weight of the new discovery reward is positive, with the weight in the early stage being less than that in the later stage. The weights of the out-of-bounds penalty and collision penalty are fixed at 1. The weight of the efficiency penalty is negative, with the weight in the early stage being greater than that in the later stage.

[0014] The performance metrics include coverage, path steps, and acquisition efficiency.

[0015] A reinforcement learning-based UAV indoor 3D reconstruction autonomous acquisition system for implementing the method, the system comprising: Scenario simulation modeling and spatial target point setting module: used to build virtual indoor simulation space in the Rhino environment and virtual drone model through the Grasshopper platform, and to set up target collection points in combination with Grasshopper scripts; The strategy learning and capability training module is used to build and train reinforcement learning models, and to design a bidirectional interface between the reinforcement learning model and the Grasshopper platform for data transmission between the two parties. The reward function of the reinforcement learning model includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. During training, at each time step, the UAV acquires target collection point information and inputs it into the reinforcement learning model. The reinforcement learning model generates action commands and transmits them to the virtual UAV model. The virtual UAV performs geometric transformations, updates its position and orientation, and reconstructs its field of view information. Autonomous Flight Evaluation and Data Acquisition Implementation Module: This module is used to deploy trained reinforcement learning models in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and to quantitatively evaluate the acquisition performance through multiple indicators.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes an autonomous data acquisition method for UAV-based 3D reconstruction of building interiors using reinforcement learning, enabling efficient and stable data acquisition in complex environments. The invention constructs multiple types of virtual indoor scenes and virtual UAV models, designs dynamically adjustable acquisition points, provides a structured spatial environment and evaluation benchmark, and selects the Proximal Policy Optimization (PPO) algorithm as the core reinforcement learning framework. A composite reward function is designed, combining factors such as coverage, efficiency, and collision penalties with dynamic weight adjustment to achieve a dual-objective optimization of "acquisition integrity and path efficiency." This mechanism effectively avoids the problems of low acquisition efficiency and high cost associated with traditional methods. Through continuous learning and adaptation of the reinforcement learning model, it maintains good performance even when the environment changes. This adaptive capability enables the UAV to effectively perform 3D reconstruction in various indoor scenes. This invention innovatively and systematically constructs a bidirectional communication interface based on GH-Python, realizing a real-time closed loop of perception input, policy generation, behavior evaluation, and action output. This achieves efficient linkage between the model and the algorithm, breaking through the limitation that the Grasshopper environment typically only supports static modeling. It allows the model to efficiently link geometric modeling and policy training, significantly improving experimental iteration efficiency. In the test experiment, the time taken for the drone to autonomously collect data for each group of rooms was 10-20 minutes, and the collection efficiency and completeness were significantly improved compared with traditional path planning. This invention has outstanding innovations in algorithm design, interface mechanism and path optimization, providing an efficient and reliable solution for 3D reconstruction in complex indoor environments such as urban renewal projects. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2This is a simulation indoor model diagram from one embodiment; Figure 3 This is a schematic diagram of the target acquisition point layout in one embodiment; Figure 4 This is a schematic diagram of the control process of the present invention; Figure 5 This is a schematic diagram of the state information in the reinforcement learning model of this invention; Figure 6 This is a schematic diagram of the reinforcement learning model training process of the present invention; Figure 7 This is a sample strategy optimization curve and fitting slope diagram of the present invention in one embodiment. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0020] This embodiment provides a method for autonomous data acquisition in indoor 3D reconstruction using unmanned aerial vehicles (UAVs) based on reinforcement learning, such as... Figure 1 As shown, the method includes the following steps: S1, Scenario Simulation Modeling and Spatial Target Point Setting: Construct a virtual indoor simulation space and a virtual drone model, and set up target acquisition points.

[0021] S11, construct a virtual indoor simulation space.

[0022] A virtual indoor simulation space was constructed using the Rhino modeling platform to simulate typical rooms, corridors, and corner nodes in a building. During the modeling process, basic spatial units were first established, and Boolean operations were used to implement openings, connections, and barriers. These units were then combined geometrically to form a controllable and complex overall environment. Subsequently, to improve realism, furniture and obstacles were introduced, and complex components were simplified by abstracting them into equivalent geometric shapes such as cuboids and cylinders, thus balancing computational efficiency with the representation of environmental features.

[0023] In the process of modeling a virtual indoor simulation space, the difficulty of model construction is adjusted according to three dimensions, forming multi-path planning and obstacle avoidance difficulty, such as... Figure 2 As shown: The difficulty is adjusted according to three dimensions: number of rooms, functional combination, and boundary obstacles. For the number of rooms, the difficulty increases from single room to multi-room combination; for functional combination, diverse spaces are formed by connecting corridors and corners; for boundary obstacles, the complexity of occlusion and obstacle avoidance is increased by adding furniture or obstacles at walls or nodes.

[0024] S12, Deploy target data collection points.

[0025] Indoor data collection points were deployed using Grasshopper scripts. Exposed surfaces of various components within the room (walls, ceiling, floor, furniture surfaces) were selected as the data collection targets, and orthogonal point clouds were deployed at 600mm intervals. Figure 3 As shown. To control training complexity, the pixel spacing and total number of pixels can be dynamically adjusted.

[0026] S13, Build a virtual drone model.

[0027] Virtual drone models can be built using the Grasshopper platform, possessing the basic characteristics of typical drones and capable of simulating flight, environmental perception, and data acquisition tasks.

[0028] S2, Policy Learning and Capability Training: A reinforcement learning model is built and trained. During training, at each time step, the UAV acquires target data point information and inputs it into the reinforcement learning model. The model generates action commands and transmits them to the bidirectional communication interface in the Grasshopper environment. After parsing, the commands are passed to the virtual UAV model, which performs geometric transformations, updates its position and orientation, and reconstructs its field of view. This information is then returned to the Python control panel, driving the training iterations. Figure 4As shown.

[0029] In reinforcement learning models, the environment state at each time step is represented as a multi-dimensional vector, such as... Figure 5 As shown, it includes: (1) Current position of the UAV; (2) Current orientation angle (one of the eight directions); (3) Number of visible target collection points in the simulated field of view; (4) Coverage (the percentage of the number of visible target collection points in the simulated field of view to the total number of target collection points); (5) Distance of the UAV to the nearest obstacle in the four directions of front, back, left, and right.

[0030] The actions output by reinforcement learning models include: (1) Translational movement: Perform a translational movement of 600mm in any of the following directions: forward, backward, left, or right; (2) Rotation action: Perform a horizontal rotation of the UAV in multiples of 45 degrees; (3) Stay in place: Remain in place for one cycle, observe or wait for a new strategy; Only one action is allowed to be executed at each time step, ensuring the interpretability and computational efficiency of trajectory control.

[0031] In the reinforcement learning training process, this invention designs a composite reward function based on multi-objective optimization theory and a dynamic weight adjustment mechanism. This function includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. The overall reward is obtained through weighted summation to comprehensively guide the UAV's path planning, expressed as: , in, For the first A reward or punishment at any time The instantaneous value, For the first A reward or punishment at any time The weight, n The weights for the number of items to be rewarded or penalized are dynamically adjusted in the early and late stages of the training phase, so that the strategy focuses on different aspects at different stages. This achieves a dynamic balance between coverage, efficiency and safety, thereby significantly improving the overall completion and training effect of UAV path planning tasks.

[0032] The coverage reward is defined as a preset value multiplied by the percentage of newly visible target collection points within the simulated field of view relative to the total number of target collection points. The new discovery reward is defined as a preset value multiplied by the number of newly discovered target collection points. The exploration reward is defined as a tiered reward, with a first score assigned to the first visit, a second score assigned to the second visit, and no further scores assigned to subsequent visits, where the first score is greater than the second score. The out-of-bounds penalty is defined as a preset penalty value assigned when the drone goes out of bounds. The collision penalty is defined as a gradient penalty based on the distance between the drone and the obstacle, with a larger penalty value for closer distances. The efficiency penalty is defined as the number of repeated visits. The coverage reward and exploration reward have positive weights, with earlier weights being greater than later weights. The new discovery reward has a positive weight, with earlier weights being less than later weights. The out-of-bounds penalty and collision penalty have a fixed weight of 1. The efficiency penalty has a negative weight, with earlier weights being greater than later weights.

[0033] In one preferred embodiment, the specific settings are shown in Table 1.

[0034] Table 1

[0035] The reinforcement learning model is trained using the Proximal Policy Optimization (PPO) algorithm as the core framework. The training process is designed as follows: Figure 6 As shown: (1) Initial environment generation: Randomly set the starting position and orientation of the drone and load the indoor target point layout information; (2) State input extraction: The Grasshopper control terminal extracts the current state and transmits it to the PPO module in Python; (3) Policy implementation and action output: The PPO network outputs actions to indicate the next movement or rotation of the drone; (4) Action execution and feedback acquisition: Grasshopper receives the action output of the execution algorithm, updates the environment, and records the changes and trajectories of the collection points; (5) Reward calculation and network update: Calculate the real-time reward based on the collection effect and use it for PPO policy update; (6) Policy optimization loop: Perform multiple rounds of training until the policy converges. By integrating the data obtained from multiple training sessions in the same room, it can be found that the collection efficiency increases with the number of training sessions, such as Figure 7 .

[0036] S3, Autonomous Flight Evaluation and Data Acquisition Implementation: Deploy a trained reinforcement learning model in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and quantitatively evaluate the acquisition performance through multiple performance indicators.

[0037] S31, Simulated Flight Experiment Setup.

[0038] The experimental environment was based on an indoor space model built with Rhino, and the training process was followed. The starting point and direction were initialized, and the target point cloud was read. Based on the current state, the policy network was invoked to output action commands. Grasshopper executed the actions (movement or rotation) and updated the spatial position. It was determined whether there were any new data collection points in the current field of view, and the data collection status was recorded. This process was repeated until the coverage target or step limit was achieved.

[0039] S32, performance evaluation.

[0040] Conduct independent experiments in the set scenarios and record key performance indicators, including coverage, path steps, and acquisition efficiency.

[0041] Based on performance metrics, the S33 will undergo a new round of adjustments and training.

[0042] This embodiment also provides an autonomous acquisition system for indoor 3D reconstruction using a UAV based on reinforcement learning, for implementing the method described above. The system includes: Scenario simulation modeling and spatial target point setting module: used to build virtual indoor simulation space in the Rhino environment and virtual drone model through the Grasshopper platform, and to set up target collection points in combination with Grasshopper scripts; The strategy learning and capability training module is used to build and train reinforcement learning models, and to design a bidirectional interface between the GH-Python-based reinforcement learning model and the Grasshopper platform, supporting state transfer, action execution, and environmental feedback. The reward function of the reinforcement learning model includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. During training, at each time step, the UAV acquires target data collection point information and inputs it into the reinforcement learning model. The reinforcement learning model generates action commands and transmits them to the virtual UAV model. The virtual UAV performs geometric transformations, updates its position and orientation, and reconstructs its field of view information. Autonomous Flight Evaluation and Data Acquisition Implementation Module: This module is used to deploy trained reinforcement learning models in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and to quantitatively evaluate the acquisition performance through multiple indicators.

[0043] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for autonomous acquisition of indoor 3D reconstruction data from unmanned aerial vehicles (UAVs) based on reinforcement learning, characterized in that, The method includes the following steps: S1, Scenario simulation modeling and spatial target point setting: Construct a virtual indoor simulation space and a virtual drone model, and set up target acquisition points; S2, Strategy Learning and Capability Training: Construct and train a reinforcement learning model. The reward function of the reinforcement learning model includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. The coverage reward is defined as a preset value multiplied by the percentage of newly visible target collection points within the simulated field of view relative to the total number of target collection points. The new discovery reward is defined as a preset value multiplied by the number of newly discovered target collection points. The exploration reward is defined as a tiered reward, assigning a first score for the first visit, a second score for the second visit, and no further scores for subsequent visits, where the first score is greater than the second score. The out-of-bounds penalty is defined as assigning a preset penalty value when the drone goes out of bounds. The collision penalty is defined as a gradient penalty based on the distance between the drone and the obstacle, with a larger penalty value for closer distances. The efficiency penalty is defined as the number of repeated visits. During training, at each time step, the drone acquires target collection point information and inputs it into the reinforcement learning model. The reinforcement learning model generates action commands and transmits them to the virtual drone model. The virtual drone performs geometric transformations, updates its position and orientation, and reconstructs the field of view information. S3, Autonomous Flight Evaluation and Data Acquisition Implementation: Deploy a trained reinforcement learning model in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and quantitatively evaluate the acquisition performance through multiple performance indicators.

2. The method for autonomous acquisition of indoor 3D reconstruction by unmanned aerial vehicles based on reinforcement learning according to claim 1, characterized in that, The construction of the virtual indoor simulation space specifically refers to: The Rhino modeling platform was used to construct a virtual indoor simulation space to simulate typical rooms, corridors and corner nodes in a building. In the modeling process, basic spatial units were first established, and openings, connections and barriers were realized through Boolean operations. The units were then combined in a geometric splicing manner to form the overall environment. Subsequently, furniture and obstacles were introduced, and complex components were simplified by abstracting them into equivalent geometric shapes.

3. The method for autonomous acquisition of indoor 3D reconstruction by UAV based on reinforcement learning according to claim 2, characterized in that, In the process of modeling the virtual indoor simulation space, the difficulty is adjusted according to three dimensions: number of rooms, functional combination, and boundary obstacles. Specifically, for the number of rooms, the difficulty increases from a single room to a combination of multiple rooms; for functional combination, diverse spaces are formed by connecting corridors and corners; and for boundary obstacles, the complexity of occlusion and obstacle avoidance is increased by adding furniture or obstacles to walls or nodes.

4. The method for autonomous acquisition of indoor 3D reconstruction by UAV based on reinforcement learning according to claim 1, characterized in that, The method for setting up target acquisition points is as follows: deployment is carried out based on Grasshopper scripts, the exposed surfaces of each component in the simulation space are selected as acquisition targets, and orthogonal point clouds are deployed at preset intervals, wherein the preset intervals are dynamically adjusted according to the complexity.

5. The method for autonomous acquisition of indoor 3D reconstruction by UAV based on reinforcement learning according to claim 1, characterized in that, In the reinforcement learning model, the environmental state at each time step is represented as a multi-dimensional vector, including the current position of the UAV, the current orientation angle, the number of visible target collection points in the simulated field of view, the coverage rate, and the distance of the UAV to the nearest obstacle in the four directions of front, back, left, and right. The coverage rate is defined as the proportion of the number of visible target collection points in the simulated field of view to the total number of target collection points.

6. The method for autonomous acquisition of indoor 3D reconstruction by UAV based on reinforcement learning according to claim 1, characterized in that, The actions output by the reinforcement learning model include: Translational movement: Perform a translation of the UAV in any of the forward, backward, left, or right directions with a preset fixed step length; Rotation action: Perform horizontal rotations of the drone by multiples of 45 degrees; Stay put: Remain in place for one cycle, observe or wait for a new strategy; Only one action is allowed to be executed at each time step.

7. The method for autonomous acquisition of indoor 3D reconstruction by UAV based on reinforcement learning according to claim 1, characterized in that, The reward function of the reinforcement learning model is expressed as: ， in, For the first A reward or punishment at any time The instantaneous value, For the first A reward or punishment at any time The weight, n The weights are dynamically adjusted as the training phase progresses, representing the number of items that will be rewarded or penalized.

8. The method for autonomous acquisition of indoor 3D reconstruction by unmanned aerial vehicles based on reinforcement learning according to claim 7, characterized in that, The weights of the coverage reward and exploration reward are positive, with the weight in the early stage being greater than that in the later stage. The weight of the new discovery reward is positive, with the weight in the early stage being less than that in the later stage. The weights of the out-of-bounds penalty and collision penalty are fixed at 1. The weight of the efficiency penalty is negative, with the weight in the early stage being greater than that in the later stage.

9. The method for autonomous acquisition of indoor 3D reconstruction by unmanned aerial vehicles based on reinforcement learning according to claim 1, characterized in that, The performance metrics include coverage, path steps, and acquisition efficiency.

10. An autonomous data acquisition system for indoor 3D reconstruction using unmanned aerial vehicles (UAVs) based on reinforcement learning, characterized in that, For implementing the method as described in any one of claims 1-9, the system comprises: Scenario simulation modeling and spatial target point setting module: used to build virtual indoor simulation space in the Rhino environment and virtual drone model through the Grasshopper platform, and to set up target collection points in combination with Grasshopper scripts; The strategy learning and capability training module is used to build and train reinforcement learning models, and to design a bidirectional interface between the reinforcement learning model and the Grasshopper platform for data transmission between the two parties. The reward function of the reinforcement learning model includes coverage reward, new discovery reward, exploration reward, out-of-bounds penalty, collision penalty, and efficiency penalty. During training, at each time step, the UAV acquires target collection point information and inputs it into the reinforcement learning model. The reinforcement learning model generates action commands and transmits them to the virtual UAV model. The virtual UAV performs geometric transformations, updates its position and orientation, and reconstructs its field of view information. Autonomous Flight Evaluation and Data Acquisition Implementation Module: This module is used to deploy trained reinforcement learning models in an untrained virtual simulation environment to conduct autonomous flight experiments and data acquisition for UAVs, and to quantitatively evaluate the acquisition performance through multiple indicators.

Images