Unmanned aerial vehicle intelligent path planning method and system based on visual signals
By optimizing the neural network using visual signals and progressive pruning techniques, a 3D semantic map is constructed. Combined with reinforcement learning for UAV path planning, the problem of localization and autonomous decision-making in complex environments for traditional UAVs is solved, achieving efficient and panoramic data acquisition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SKILL TRAINING CENT STATE GRID JIBEI ELECTRONICS POWER COMPANY
- Filing Date
- 2026-01-28
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional UAV navigation systems are inaccurate in positioning in complex environments and lack autonomous decision-making capabilities. Existing path planning algorithms lack multi-dimensional target optimization and are unable to achieve high-quality data collection for key targets.
We employ a visual signal-based intelligent path planning method for unmanned aerial vehicles (UAVs). By progressively pruning and optimizing the neural network, we construct a three-dimensional semantic map and combine reinforcement learning for real-time path planning and autonomous operation, thereby achieving a balance between multiple objectives.
Achieve precise navigation in complex environments, reduce computational complexity, autonomously plan optimal observation paths, realize panoramic data acquisition, and reduce reliance on human intervention.
Smart Images
Figure CN121857771A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of path planning technology, specifically relating to a method and system for intelligent path planning of unmanned aerial vehicles (UAVs) based on visual signals. Background Technology
[0002] With the rapid development and widespread application of drone technology, its value in fields such as power line inspection, disaster relief, and environmental monitoring is becoming increasingly prominent. Traditional drone navigation systems mainly rely on satellite positioning technologies such as GPS / RTK. However, in practical application scenarios, especially in complex environments such as mountainous areas, urban canyons, indoor spaces, or disaster sites, these positioning signals are often severely interfered with or even completely fail, making it difficult for drones to accurately locate and navigate. At the same time, existing drone systems have significant shortcomings in autonomous decision-making capabilities, mostly relying on preset routes or manual control, unable to flexibly adjust flight strategies according to real-time environmental changes. Although vision-based navigation technologies (such as visual SLAM and optical flow navigation) have made some progress, these methods often have high computational complexity and insufficient robustness, making it difficult to achieve real-time stable operation on drone platforms with limited computing resources. Furthermore, existing path planning algorithms typically only focus on simple obstacle avoidance or shortest paths, lacking comprehensive optimization of multi-dimensional objectives (such as observation quality, task completion, and energy efficiency), resulting in poor performance in complex task scenarios. Especially in the autonomous operation phase, traditional systems struggle to automatically plan the optimal observation path based on target characteristics, failing to achieve comprehensive, high-quality data acquisition of key targets. Summary of the Invention
[0003] To address the aforementioned problems in the existing technology, this invention provides a method and system for intelligent path planning of unmanned aerial vehicles (UAVs) based on visual signals.
[0004] The objective of this invention can be achieved through the following technical solutions: A visual signal-based intelligent path planning method for unmanned aerial vehicles (UAVs) includes the following steps: Step S1: Optimize the neural network through progressive pruning; Step S2: Provide the drone with fuzzy coordinates and scene type. The drone flies to the fuzzy coordinate area. When it approaches the fuzzy coordinate area, it switches to pure vision mode to perceive the environment through the neural network and construct a three-dimensional semantic map. Step S3: Perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; Step S4: Fly autonomously along the flight path. When approaching the target, perform autonomous operations, obtain real-time monitoring data, store it in the onboard storage device, and transmit it back after returning to base.
[0005] Preferably, the progressive pruning in step S1 specifically includes: A representative image set is preset. Images in the representative image set are processed by a neural network, and the output value of each image is recorded. For each neuron, the average value of its output value across the entire representative image set is obtained and denoted as the average activation intensity of the neuron. The connection weights between two neurons are obtained, and the connection importance is derived from the average activation strength and the connection weights. Mathematically, this is described as follows: ,in, To illustrate the importance of connections between neurons i and j in layer r, The maximum attenuation percentage for each round of pruning. For the current pruning cycle, The number of rounds required to achieve maximum decay Let be the connection weights of neurons i and j in the r-th layer. Let i be the average activation intensity of neuron i in layer r. represents the average activation intensity of neuron j in the (r-1)th layer; All connections are sorted according to their importance, a pruning ratio is determined, and connections with low importance are pruned according to the pruning ratio. After multiple rounds of pruning, the optimized neural network is output.
[0006] Preferably, the environmental perception in step S2 specifically includes: Real-time acquisition of video stream, outputting raw image frame sequence {I1,I2,…,I...} t}; The current image frame I is processed using a convolutional neural network. t Perform object detection and output the bounding boxes, category labels, and confidence scores of all objects of interest in the current image frame; The current image frame I is analyzed using a monocular depth estimation network. t Perform depth estimation and output the result relative to the current image frame I. t Depth map of the same size; For each bounding box, the covered pixel region is selected, the depth value of the corresponding region in the depth map is queried, and the statistical value of the depth value is calculated as the representative depth of the corresponding object; the center pixel coordinates of the bounding box are obtained, and the center pixel coordinates and the representative depth are converted into three-dimensional coordinates (x, y, y) relative to the UAV coordinate system. i ,y i ,z i ), and assign the category label c to that point. i ; The current image frame I t The three-dimensional semantic point p i = (xi ,y i ,z i ,c i The three-dimensional semantic map is formed by fusing the three-dimensional semantic points of the previous image frame with the three-dimensional semantic points of the previous image frame.
[0007] Preferably, the real-time path planning in step S3 specifically includes: Obtain the current status and target orientation of the drone; The current state of the UAV and the target direction are encoded into a state vector; The state vector is input into the reinforcement learning policy network, and the action vector is output. A set of candidate actions is generated based on the state vector, the action vector, and the new state vector reached after executing the action vector. For each candidate action, a multi-objective reward function is constructed, mathematically described as follows: ,in, For instant rewards, As a schedule weight, For safety weights, For observation weights, As energy consumption weight, As a progress reward, As a safety reward, As a reward for observation, As an energy consumption reward; The flight trajectory is obtained based on the instant reward.
[0008] Preferably, the autonomous operation in step S4 specifically includes: The system continuously acquires current image frames, analyzes the current image frames using a semantic segmentation network, and combines the three-dimensional semantic map to locate the task target and obtain the three-dimensional bounding box of the task target. Define a set of candidate observation perspectives, each of which corresponds to a candidate inspection waypoint; Based on the aforementioned 3D bounding box, the current pose of the UAV, and its observation history, an evaluation function is constructed to select inspection waypoints from the candidate inspection waypoints. Mathematically, this is described as follows: ,in, Let v be the reward score for the candidate inspection waypoint. For observation weights, For intersection, union, and comparison, This represents the visible portion of the target T within the camera's field of view when observing the target T from a candidate inspection waypoint v. For an ideal viewing view, To cover weights, The historical coverage of the target being operated on is the percentage of time it has been observed. Distance weights From candidate inspection waypoint v to current location distance, The maximum allowed waypoint distance; The candidate waypoint with the highest benefit score is selected as the next waypoint to be inspected. The drone performs real-time monitoring according to the inspection route points and collects the real-time monitoring data.
[0009] A visual signal-based UAV intelligent path planning system is used to execute the aforementioned visual signal-based UAV intelligent path planning method, including a progressive pruning module, a map building module, a path planning module, and an autonomous operation module. The progressive pruning module is used to optimize the neural network through progressive pruning; The map building module is used to provide the drone with fuzzy coordinates and scene type. The drone flies to the fuzzy coordinate area. When it approaches the fuzzy coordinate area, it switches to pure vision mode to perceive the environment through the neural network and build a three-dimensional semantic map. The path planning module is used to perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; The autonomous operation module is used to fly autonomously along the flight trajectory. When it approaches the target, it performs autonomous operations, obtains real-time monitoring data, stores it in the onboard storage device, and transmits it back after returning to base.
[0010] The beneficial effects of this invention are as follows: (1) By replacing or enhancing traditional satellite positioning with visual signals, UAVs can maintain accurate navigation in environments where GPS / RTK signals are weak or completely lost, greatly expanding the application boundaries of UAVs, especially suitable for scenarios that are difficult to cover by traditional navigation technologies, such as disaster sites, mountain valleys, and indoor spaces.
[0011] (2) By using progressive pruning techniques, computational complexity and memory usage are significantly reduced while maintaining network performance, enabling complex deep learning models to run efficiently on resource-constrained UAV platforms, achieving the best balance between algorithm performance and hardware limitations.
[0012] (3) By integrating multiple dimensions of objectives such as progress, safety, observation quality and energy efficiency into a reinforcement learning framework, the UAV can autonomously weigh various factors according to the real-time environment and mission requirements to generate the optimal flight strategy, which not only ensures mission completion efficiency, but also takes into account safety and endurance.
[0013] (4) Through the autonomous target recognition and multi-view planning algorithm based on the three-dimensional semantic map, the UAV can intelligently identify key targets and automatically plan the optimal observation path, realize panoramic and blind-spot-free data collection of targets, greatly improve the quality and efficiency of inspection, and reduce the dependence on manual intervention. Attached Figure Description
[0014] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.
[0015] Figure 1 This is a flowchart illustrating the steps of a visual signal-based intelligent path planning method for unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation
[0016] To better understand the invention, various aspects of the invention will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely illustrative of exemplary embodiments of the invention and are not intended to limit the scope of the invention in any way. Throughout the specification, the expression "and / or" includes any and all combinations of one or more of the associated listed items. As used herein, the terms "approximately," "about," and similar terms are used as expressions of approximation, not as expressions of degree, and are intended to describe inherent deviations in measured or calculated values that will be recognized by those skilled in the art. Furthermore, the order in which the steps are described in this invention does not necessarily indicate the order in which these steps occur in actual operation, unless otherwise expressly defined or deduced from the context.
[0017] It should also be understood that expressions such as "comprising," "including," "having," "containing," and / or "comprising" are open-ended rather than closed-ended expressions in this specification, indicating the presence of the stated features, elements, and / or components, but not excluding the presence of one or more other features, elements, components, and / or combinations thereof. Furthermore, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features, not just individual elements in the list. Additionally, when describing embodiments of the invention, the word "may" is used to mean "one or more embodiments of the invention." And the term "exemplary" is intended to refer to examples or illustrations.
[0018] Unless otherwise specified, all terms used herein (including engineering and technical terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that, unless expressly stated herein, terms defined in common dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the relevant art, and not in an idealized or overly formalized sense.
[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0020] Example 1: Please see Figure 1 A visual signal-based intelligent path planning method for unmanned aerial vehicles (UAVs) includes: Step S1: Optimize all neural networks required for subsequent steps through progressive pruning; Step S2: The command center receives a disaster alert and provides the drone with fuzzy coordinates (such as latitude and longitude range) and scene type (such as "landslide caused line damage"). The drone takes off from the base and uses GPS / RTK navigation to fly to the fuzzy coordinate area. When it approaches the fuzzy coordinate area or detects that the GPS signal is weak, the drone switches to pure vision mode to perceive the environment through the neural network and build a three-dimensional semantic map. Step S3: Perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; Step S4: The UAV flies autonomously along the flight path and continues to perform the above steps, making real-time fine adjustments to the trajectory. When it approaches the target, the UAV performs autonomous operations, obtains real-time monitoring data, stores it in the onboard storage device, and transmits it back after returning to base.
[0021] In this embodiment, the progressive pruning specifically refers to the following: A neural network consists of a large number of neurons and the connections between them. Each connection has a weight value, which determines the importance of a signal transmitted from one neuron to another. A well-trained neural network has a large number of connections, but in actual operation, the weight values of many connections are close to 0, meaning that these connections contribute almost nothing to the network's output while occupying a large amount of computing resources and memory space. In the process of intelligent path planning for unmanned aerial vehicles (UAVs), a large number of neural network models are required. Therefore, we need to use progressive pruning to remove unimportant connections in the neural network to maintain good network performance. S101: A representative image set (e.g., several thousand images of power equipment) is preset. The images in the representative image set are processed by a neural network and the output value of each image is recorded. For each neuron, the average value of its output value on the entire representative image set is obtained and recorded as the average activation intensity of the neuron. S102: Obtain the connection weights between two neurons (fixed after training). Calculate the connection importance using the average activation strength and the connection weights, mathematically described as follows: ,in, To illustrate the importance of connections between neurons i and j in layer r, The maximum attenuation percentage for each round of pruning. For the current pruning cycle, The number of rounds required to achieve maximum decay Let be the connection weights of neurons i and j in the r-th layer. Let i be the average activation intensity of neuron i in layer r. represents the average activation intensity of neuron j in the (r-1)th layer; S103: Sort all connections according to their importance, determine the pruning ratio, prune connections with low importance according to the pruning ratio, and output the optimized neural network after multiple rounds of pruning.
[0022] In this embodiment, the environmental perception specifically refers to: S201: Real-time acquisition of video streams via the UAV's onboard camera, outputting the raw image frame sequence {I1,I2,…,I…} t}; S202: Process the current image frame I using a lightweight convolutional neural network. t Perform target detection and output the bounding boxes, category labels (such as "power tower", "wire", "insulator", "tree", "house" etc.) and confidence scores of all objects of interest in the current image frame; S203: The current image frame I is processed by a lightweight monocular depth estimation network. t Perform depth estimation and output the result relative to the current image frame I. t A depth map of the same size, wherein each pixel value in the depth map represents the estimated distance between that point and the camera; S204: For each bounding box, take the pixel region it covers, query the depth value of the corresponding region in the depth map, and calculate the statistic of the depth value (such as the median) as the representative depth of the corresponding object; obtain the center pixel coordinates of the bounding box, and convert the center pixel coordinates and the representative depth into three-dimensional coordinates (x, y, y) relative to the UAV coordinate system. i ,y i ,z i ), and assign the category label c to that point. i ; S205: Set the current image frame I t The three-dimensional semantic point p i = (x i ,y i ,z i ,c i The three-dimensional semantic points of the image are fused with those of the previous image frames to form the global three-dimensional semantic map.
[0023] In this embodiment, the real-time path planning specifically includes: S301: Obtain the current state of the drone (position, speed, battery level, etc.) and target direction (one or more regions of interest inferred from fuzzy coordinates and scene type); S302: Encode the current state of the UAV and the target direction into a fixed-dimensional state vector; S303: Input the state vector into the trained reinforcement learning policy network (a deep neural network) and output the action probability distribution or deterministic action vector, such as the speed increment and yaw angle change of the UAV in the next time step. S304: Generate a set of candidate actions based on the state vector, the action vector, and the new state vector reached after executing the action vector; S305: For each candidate action, construct a multi-objective reward function, mathematically described as follows: ,in, For instant rewards, As a schedule weight, For safety weights, For observation weights, As energy consumption weight, As a progress reward, As a safety reward, As a reward for observation, As an energy consumption reward, the progress reward is based on the change in distance the UAV travels towards the target region of interest, let d t Let t be the shortest distance from the UAV to the target area (considering obstacles), then =d t -d t+1 That is, the reduction in distance is a positive reward; the safety reward is based on the minimum distance between the drone and the obstacle, let d min The minimum distance between the drone and all obstacles (points labeled as obstacles in the 3D semantic map) is defined as follows: if it is less than the safety threshold d... safe ,but Otherwise, it is 0, and C is the penalty coefficient; the observation reward encourages the UAV to adjust its attitude to better observe the target, obtain the pixel area and position of the target object in the image in the current frame, define a function, and the reward is high when the target is in the center of the image and the area is moderate. At the same time, considering whether the viewpoint is frontal, it can be evaluated by the angle between the normal vector of the target object's three-dimensional point cloud and the camera's optical axis; the energy consumption reward is negatively correlated with the amplitude of the action. S306: The flight trajectory is obtained based on the instant reward (i.e., the candidate action with the highest instant reward is selected at each step, and finally a flight trajectory is formed by these candidate actions).
[0024] In this embodiment, the autonomous operation specifically refers to: S401: When the UAV is in an area with no or weak signal, in addition to basic navigation and obstacle avoidance, it also needs to autonomously decide which operational targets to inspect in detail and automatically generate inspection waypoints; continuously collect the current image frame, use a lightweight semantic segmentation network to analyze the current image frame, combine it with the three-dimensional semantic map, lock the operational targets that need to be inspected in detail (such as a collapsed power tower), and obtain the three-dimensional bounding box of the operational target. S402: In order to obtain comprehensive information about the target, it is necessary to operate the UAV to observe from multiple perspectives and define a set of candidate observation perspectives (for example, a point every 30 degrees in the horizontal circle around the target, and 30 degrees up and down in the vertical direction, which includes not only the position, but also the orientation and other pose information of the UAV). Each candidate observation perspective corresponds to a candidate inspection waypoint. S403: Based on the aforementioned 3D bounding box, the current pose of the UAV, and the observation history, an evaluation function is constructed to select inspection waypoints from the candidate inspection waypoints. Mathematically, this is described as follows: ,in, Let v be the reward score for the candidate inspection waypoint. For observation weights, For intersection, union, and comparison, This represents the visible portion of the target T in the camera's field of view when observing the target T from the candidate inspection waypoint v (the projection area of the 3D bounding box in the current image frame, simulated geometrically). For an ideal viewing view, such as an image region where the target is complete and centered, To cover weights, The historical coverage of the target has been observed (calculate the proportion of the target surface that has been observed by past waypoints). The program encourages drones to explore unobserved target areas. Once a drone detects a target, it first flies to its front and takes a clear picture of the front. At this point, the COV value on the back is very low, as shown in the formula. The system will guide the drone to fly around to the back and take pictures, thereby automatically achieving panoramic, no-blind-spot data collection of the target. Distance weights From candidate inspection waypoint v to current location distance, The maximum allowable waypoint distance is determined by selecting the candidate waypoint with the highest benefit score as the next waypoint to be inspected. S404: The UAV performs real-time monitoring according to the inspection waypoints and collects the real-time monitoring data.
[0025] Example 2: A visual signal-based intelligent path planning system for unmanned aerial vehicles (UAVs) includes a progressive pruning module, a map building module, a path planning module, and an autonomous operation module. The progressive pruning module is used to optimize all the neural networks required in subsequent steps through progressive pruning; The map building module is used by the command center to receive disaster alarms, provide fuzzy coordinates and scene types for drones, and the drones take off from the base and use GPS / RTK navigation to fly to the fuzzy coordinate area. When approaching the fuzzy coordinate area or detecting a weakening GPS signal, the drones switch to pure vision mode to perceive the environment through the neural network and build a three-dimensional semantic map. The path planning module is used to perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; The autonomous operation module is used for the UAV to fly autonomously along the flight path and continuously perform the above steps, fine-tune the trajectory in real time, and when it approaches the operation target, the UAV performs autonomous operation, obtains real-time monitoring data, stores it in the onboard storage device, and transmits it back after returning to base.
[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for intelligent path planning of unmanned aerial vehicles (UAVs) based on visual signals, characterized in that, Includes the following steps: Step S1: Optimize the neural network through progressive pruning; Step S2: Provide the drone with fuzzy coordinates and scene type. The drone flies to the fuzzy coordinate area. When it approaches the fuzzy coordinate area, it switches to pure vision mode to perceive the environment through the neural network and construct a three-dimensional semantic map. Step S3: Perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; Step S4: Fly autonomously along the flight path. When approaching the target, perform autonomous operations, obtain real-time monitoring data, store it in the onboard storage device, and transmit it back after returning to base.
2. The intelligent path planning method for unmanned aerial vehicles based on visual signals according to claim 1, characterized in that, The progressive pruning in step S1 specifically refers to: A representative image set is preset. Images in the representative image set are processed by a neural network, and the output value of each image is recorded. For each neuron, the average value of its output value across the entire representative image set is obtained and denoted as the average activation intensity of the neuron. The connection weights between two neurons are obtained, and the connection importance is derived from the average activation strength and the connection weights. Mathematically, this is described as follows: ,in, To illustrate the importance of connections between neurons i and j in layer r, The maximum attenuation percentage for each round of pruning. For the current pruning cycle, The number of rounds required to achieve maximum decay Let be the connection weights of neurons i and j in the r-th layer. Let i be the average activation intensity of neuron i in layer r. represents the average activation intensity of neuron j in the (r-1)th layer; All connections are sorted according to their importance, a pruning ratio is determined, and connections with low importance are pruned according to the pruning ratio. After multiple rounds of pruning, the optimized neural network is output.
3. The intelligent path planning method for unmanned aerial vehicles based on visual signals according to claim 1, characterized in that, The environmental perception mentioned in step S2 specifically refers to: Real-time acquisition of video stream, outputting raw image frame sequence {I1,I2,…,I...} t }; The current image frame I is processed using a convolutional neural network. t Perform object detection and output the bounding boxes, category labels, and confidence scores of all objects of interest in the current image frame; The current image frame I is analyzed using a monocular depth estimation network. t Perform depth estimation and output the result relative to the current image frame I. t Depth map of the same size; For each bounding box, the covered pixel region is selected, the depth value of the corresponding region in the depth map is queried, and the statistical value of the depth value is calculated as the representative depth of the corresponding object; the center pixel coordinates of the bounding box are obtained, and the center pixel coordinates and the representative depth are converted into three-dimensional coordinates (x, y, y) relative to the UAV coordinate system. i ,y i ,z i ), and assign the category label c to that point. i ; The current image frame I t The three-dimensional semantic point p i = (x i ,y i ,z i ,c i The three-dimensional semantic map is formed by fusing the three-dimensional semantic points of the previous image frame with the three-dimensional semantic points of the previous image frame.
4. The intelligent path planning method for unmanned aerial vehicles based on visual signals according to claim 1, characterized in that, The real-time path planning in step S3 specifically refers to: Obtain the current status and target orientation of the drone; The current state of the UAV and the target direction are encoded into a state vector; The state vector is input into the reinforcement learning policy network, and the action vector is output. A set of candidate actions is generated based on the state vector, the action vector, and the new state vector reached after executing the action vector. For each candidate action, a multi-objective reward function is constructed, mathematically described as follows: ,in, For instant rewards, As a schedule weight, For safety weights, For observation weights, As energy consumption weight, As a progress reward, As a safety reward, As a reward for observation, As an energy consumption reward; The flight trajectory is obtained based on the instant reward.
5. The intelligent path planning method for unmanned aerial vehicles based on visual signals according to claim 1, characterized in that, The autonomous operation mentioned in step S4 specifically refers to: The system continuously acquires current image frames, analyzes the current image frames using a semantic segmentation network, and combines the three-dimensional semantic map to locate the task target and obtain the three-dimensional bounding box of the task target. Define a set of candidate observation perspectives, each of which corresponds to a candidate inspection waypoint; Based on the aforementioned 3D bounding box, the current pose of the UAV, and its observation history, an evaluation function is constructed to select inspection waypoints from the candidate inspection waypoints. Mathematically, this is described as follows: ,in, Let v be the reward score for the candidate inspection waypoint. For observation weights, For intersection, union, and comparison, This represents the visible portion of the target T within the camera's field of view when observing the target T from a candidate inspection waypoint v. For an ideal viewing view, To cover weights, The historical coverage of the target being operated on is the percentage of time that has been observed. Distance weights From candidate inspection waypoint v to current location distance, The maximum allowed waypoint distance; The candidate waypoint with the highest benefit score is selected as the next waypoint to be inspected. The drone performs real-time monitoring according to the inspection route points and collects the real-time monitoring data.
6. A visual signal-based intelligent path planning system for unmanned aerial vehicles (UAVs), characterized in that, The system is applied to the visual signal-based UAV intelligent path planning method as described in any one of claims 1-5, including a progressive pruning module, a map building module, a path planning module, and an autonomous operation module; The progressive pruning module is used to optimize the neural network through progressive pruning; The map building module is used to provide the drone with fuzzy coordinates and scene type. The drone flies to the fuzzy coordinate area. When it approaches the fuzzy coordinate area, it switches to pure vision mode to perceive the environment through the neural network and build a three-dimensional semantic map. The path planning module is used to perform real-time path planning based on the three-dimensional semantic map to obtain the flight trajectory; The autonomous operation module is used to fly autonomously along the flight trajectory. When it approaches the target, it performs autonomous operations, obtains real-time monitoring data, stores it in the onboard storage device, and transmits it back after returning to base.