Unmanned aerial vehicle high-altitude cleaning path precise planning method and system based on visual navigation

Abstract

The invention discloses an unmanned aerial vehicle high-altitude cleaning path precise planning method and system based on visual navigation, and the method comprises the following steps: constructing an initial three-dimensional grid model of a to-be-cleaned building, dividing cleaning regions according to the initial three-dimensional grid model, and setting priorities; the unmanned aerial vehicle collects visual data of a cleaning area through a multi-mode visual sensor, wherein the visual data comprise surface texture, depth information and a stain distribution thermodynamic diagram; performing local correction on the initial three-dimensional grid model based on the real-time visual data to generate a high-precision dynamic three-dimensional grid model; generating a cleaning path by adopting an adaptive grid traversal algorithm according to the priority of the cleaning area and the coverage range of the unmanned aerial vehicle cleaning head; obstacles in the path are detected in real time through a visual feature matching algorithm, and when the obstacles are detected, path re-planning is carried out to avoid the obstacles until all the areas are cleaned. According to the method, the modeling precision and the dynamic adaptability are improved, and the intelligence and the pertinence of path planning are realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent path planning, and particularly relates to a method and system for precise path planning of high-altitude cleaning of an unmanned aerial vehicle based on visual navigation. BACKGROUND

[0002] With the popularity of modern high-rise buildings, the demand for external wall cleaning, which is necessary for ensuring the cleanliness and structural safety of buildings, is increasing. Traditional manual high-altitude cleaning methods not only have high safety risks, but also are inefficient and costly, making it difficult to meet the large-scale and high-frequency cleaning needs. Unmanned aerial vehicle (UAV) high-altitude cleaning technology, with its flexibility, safety, and efficiency, has gradually become an important direction for the development of the industry, and precise path planning is a key link that determines the quality and efficiency of UAV cleaning operations.

[0003] Current UAV high-altitude cleaning path planning technology still has many key defects, which restrict its actual application effect: 1. Insufficient precision and dynamic adaptability of building modeling Traditional three-dimensional modeling methods for buildings rely on manual measurement, static CAD drawings, or low-precision pre-scanning data, which have the following problems: Low efficiency and safety risks: manual measurement of high-altitude building structures requires the use of scaffolding or climbing equipment, which has a long operation period and high risk coefficient; Poor model accuracy: static CAD models cannot reflect actual wear and tear, deformation, or temporary structural changes on the building surface, leading to a disconnect between path planning and the actual scene; Lack of dynamic updating capability: static models generated by pre-scanning cannot adapt to dynamic scenarios such as changes in stains on the building surface during cleaning and temporary obstacles, which can easily cause missed cleaning and collisions.

[0004] 2. Lack of targetedness and intelligence in path planning Existing path planning methods mostly use global uniform traversal strategies (such as snake-shaped and spiral-shaped), without fully optimizing them based on building features and cleaning needs: Poor regional adaptability: uniform traversal strategies cannot adapt to complex and diverse geometric features of building surfaces (such as flat surfaces, curved surfaces, and mixed structures with corners), which can easily lead to missed cleaning in curved surface areas and repeated cleaning in flat surface areas; Lack of priority logic: artificial setting of cleaning order lacks quantitative basis and often ignores the priority processing of areas with high stain density and high operation difficulty (such as high-altitude vertical surfaces and complex structural parts), resulting in unreasonable allocation of resources such as UAV power and time; Fixed path parameters: parameters such as the distance between cleaning paths and turning points are fixed values and are not dynamically adjusted based on the coverage range of the cleaning head and the distribution of stains on the building surface, which can easily result in missed cleaning or resource waste.

[0005] Three, weak obstacle detection and obstacle avoidance capability Temporary obstacles (such as scaffolding, billboards, air conditioner outdoor units) frequently occur in high-altitude environments, and the existing technology has obvious deficiencies in response capability: Single sensing means: only relying on a single sensor of RGB camera or depth camera, easy to miss transparent, low-contrast or special obstacle with thermal characteristics, high collision risk; Rigid obstacle avoidance strategy: regardless of the size or influence range of the obstacle, a unified avoidance method is adopted, which leads to low efficiency in small obstacle scenarios and unreasonable safety distance setting in large obstacle scenarios, which easily causes collision; Dynamic response lag: relying on static building models for path planning, unable to perceive and respond to sudden obstacles during cleaning in real time, easy to cause cleaning interruption or missed cleaning.

[0006] In summary, the existing unmanned aerial vehicle high-altitude cleaning path planning technology has significant shortcomings in modeling accuracy, path intelligence, dynamic obstacle avoidance capability, etc. A technical solution combining real-time visual perception, dynamic model updating and adaptive path planning is needed to improve the accuracy, safety and efficiency of unmanned aerial vehicle high-altitude cleaning operation. SUMMARY

[0007] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a method and system for precise planning of unmanned aerial vehicle high-altitude cleaning path based on visual navigation, to improve modeling accuracy and dynamic adaptability, realize the intelligence and pertinence of path planning, enhance obstacle detection and dynamic obstacle avoidance capability, and comprehensively improve the accuracy, safety and efficiency of high-altitude cleaning operation.

[0008] To solve the above problems, the technical solution adopted by the present application is as follows: A method for precise planning of unmanned aerial vehicle high-altitude cleaning path based on visual navigation, comprising the following steps: Constructing an initial three-dimensional grid model of the building to be cleaned, dividing the cleaning area according to the initial three-dimensional grid model and setting the priority; The unmanned aerial vehicle collects visual data of the cleaning area through multi-modal visual sensors, including surface texture, depth information and stain distribution thermal map; Based on real-time visual data, the initial three-dimensional grid model is locally corrected to generate a high-precision dynamic three-dimensional grid model; Using an adaptive grid traversal algorithm, a cleaning path is generated according to the priority of the cleaning area and the coverage range of the unmanned aerial vehicle cleaning head; Real-time detection of obstacles in the path is performed through a visual feature matching algorithm, and when an obstacle is detected, the path is re-planned to avoid the obstacle until all areas are cleaned.

[0009] Preferably, when constructing the initial three-dimensional mesh model of the building to be cleaned, the following steps are included: The building surface color image is collected by the RGB camera carried by the pre-flight UAV, and the distance information is collected by the depth camera. The rough three-dimensional point cloud model is generated by combining the simultaneous localization and mapping algorithm. Then, the initial three-dimensional mesh model is converted by mesh simplification and texture mapping.

[0010] Preferably, when dividing the cleaning area and setting the priority according to the initial three-dimensional mesh model, the following steps are included: Based on the surface geometric features of the initial three-dimensional mesh model, the building surface is divided into multiple independent cleaning areas; The average of the stain distribution density of each area is preliminarily evaluated by using the pre-flight infrared thermal imaging data or color image texture analysis. Combined with the vertical height and area size of the area, the priority score is calculated by using a weighted formula; According to the priority score from high to low, the cleaning priority order of each cleaning area is determined.

[0011] Preferably, when collecting the visual data of the cleaning area, the following steps are included: The color image of the cleaning area is collected by the RGB camera at a preset frame rate. The texture features of the image are extracted, and the surface texture feature map containing color distribution and texture complexity is generated; The distance data of each point in the cleaning area relative to the UAV is obtained in real time by the depth camera. Combined with the real-time pose information of the UAV, the three-dimensional space depth coordinates of each point in the cleaning area are calculated; The thermal radiation intensity data of the cleaning area is collected by the infrared thermal imaging module. The thermal radiation value is mapped into a thermal map with different color gradients. At the same time, combined with the analysis result of the proportion of dark stain area in the color image collected by the RGB camera, the thermal map is corrected to obtain a stain distribution thermal map that accurately reflects the stain distribution.

[0012] Preferably, when generating a high-precision dynamic three-dimensional mesh model, the following steps are included: The real-time collected surface texture feature map, three-dimensional space depth coordinates and stain distribution thermal map are converted to the unified coordinate system of the initial three-dimensional mesh model through the real-time pose information of the UAV, to ensure the spatial consistency of the data; The deviation of the geometric parameters of the corresponding area of the initial three-dimensional mesh model and the three-dimensional space depth coordinates data, and the detail difference between the initial texture mapping and the surface texture feature map are compared. The area with a deviation exceeding a preset threshold is selected as a to-be-corrected area; For the to-be-corrected area, the three-dimensional space depth coordinates data is used to adjust the position of the grid vertex and optimize the local grid topology to match the actual geometric shape of the building surface; The initial texture map of the to-be-corrected region is replaced by the surface texture feature map to retain the real surface details, and the high-stain density area in the stain distribution thermal map is marked as a key area of the dynamic three-dimensional grid model to strengthen the accuracy of the geometric and texture details thereof.

[0013] Preferably, when generating the cleaning path, the following steps are included: According to the cleaning priority order of each cleaning area, the current cleaning area is assigned; According to the surface geometric features of the cleaning area, the optimal traversal direction is determined; According to the coverage range of the unmanned aerial vehicle cleaning head, the spacing and length of the path segments are adjusted so that the overlap rate of adjacent path segments is controlled within 10%-15%; The path turning points are fitted by using a Bezier curve, and the path with the same direction is preferentially selected.

[0014] Preferably, when detecting obstacles in the path in real time by using a visual feature matching algorithm, the following steps are included: Local feature points are extracted from the real-time frame of the RGB camera of the unmanned aerial vehicle, and three-dimensional space depth coordinates corresponding to the local feature points are obtained from the depth camera, and thermal radiation intensity values corresponding to the local feature points are obtained from the infrared thermal imaging module; The extracted local feature points are matched with the pre-stored feature library of the corresponding area in the dynamic three-dimensional grid model, and the position deviation, depth deviation and thermal radiation deviation of the feature points are calculated; The position deviation threshold, depth deviation threshold and thermal radiation deviation threshold are preset, and when any deviation exceeds the corresponding threshold, the area where the feature point is located is marked as a suspected obstacle area; For the suspected obstacle area, after the existence of the obstacle is confirmed through continuous multi-frame visual data tracking verification, the three-dimensional space depth coordinates, the circumscribed rectangle size and the contour shape of the obstacle are calculated by using the depth camera data.

[0015] Preferably, when the path is re-planned after detecting the obstacle, the following steps are included: The three-dimensional space depth coordinates, the circumscribed rectangle size and the contour shape of the obstacle are used to determine the influence range of the obstacle in combination with the geometric boundary of the corresponding area in the dynamic three-dimensional grid model; An optimal strategy is selected according to the relative positional relationship between the influence range of the obstacle and the current cleaning path, wherein: If the influence range of the obstacle is less than 30% of the length of the current path segment, a temporary jump strategy is adopted to adjust the position of the end point of the current path segment, skip the influence range of the obstacle, and then restore the original traversal direction; If the influence range of the obstacle is greater than 30% of the length of the current path segment, a detour strategy is adopted, and the shortest detour path from the current position to the target position on the other side of the obstacle is searched in the dynamic three-dimensional grid model based on the A* algorithm or the rapidly-exploring random tree algorithm, and after the detour is completed, the original traversal direction is restored.

[0016] Preferably, in searching the shortest detour path, the following is included: Based on the three-dimensional spatial depth coordinates, the circumscribed rectangular dimensions, and the contour shape of the obstacle in the dynamic three-dimensional grid model, combined with the retained building structure boundaries in the model, it is clear that the path must not pass through the obstacle body and the building structure boundaries, and the minimum safety distance of the path from the obstacle is not less than 20% of the coverage range of the unmanned aerial vehicle cleaning head; A multi-dimensional cost function is constructed, including: the total length of the path, the safety distance of the path from the obstacle, the curvature change rate of the path turning point, and the overlap degree of the path coverage area and the high-stain density area in the stain distribution thermal map; The A* algorithm or the rapidly-exploring random tree algorithm is used to generate an initial detour path based on the multi-dimensional cost function in the search space corresponding to the dynamic three-dimensional grid model, with the current position as the starting point and the target position on the other side of the obstacle as the ending point; The initial detour path is smoothed and the path segment spacing is adjusted to ensure that the overlap rate of adjacent path segments is maintained at 10%-15%; Whether the optimized path meets the constraint conditions and the cleaning coverage requirements is verified.

[0017] A visual navigation-based unmanned aerial vehicle high-altitude cleaning path accurate planning system, when running, executes the above method, including: An unmanned aerial vehicle body carrying a flight control unit and a cleaning execution device; A multi-modal visual perception unit including an RGB camera, a depth camera, and an infrared thermal imaging module, for collecting surface texture, depth information, and stain distribution thermal map of the cleaning area; An initial model construction module for generating an initial three-dimensional grid model and dividing the cleaning area and setting priorities; A dynamic model updating module for locally correcting the initial three-dimensional grid model based on real-time visual data to generate a high-precision dynamic three-dimensional grid model; A path planning module for generating a cleaning path using an adaptive grid traversal algorithm; An obstacle detection and path re-planning module for real-time detection of obstacles in the path through a visual feature matching algorithm and path re-planning to avoid obstacles; A cleaning execution control module for controlling the unmanned aerial vehicle body to complete high-altitude cleaning operations through the cleaning execution device according to the planned path. Compared with the prior art, the beneficial effects of the present application are: (1) Modeling accuracy and dynamic adaptability improvement An initial three-dimensional grid model is generated by pre-flight of the unmanned aerial vehicle combined with the SLAM algorithm, replacing the traditional manual measurement or static CAD modeling method, avoiding the safety risks of high-altitude operations, greatly shortening the modeling cycle, and at the same time ensuring that the model can reflect the geometric shape and texture characteristics of the building surface, providing an accurate digital basis for subsequent path planning.

[0018] Based on real-time data collected by multi-modal visual sensors, the initial three-dimensional grid model is locally corrected to generate a high-precision dynamic three-dimensional grid model. It can adapt to dynamic scenarios such as wear, deformation, stain changes, and temporary structural adjustments on the building surface in a timely manner, solving the problem of disconnection between traditional static models and actual scenes, and ensuring that path planning always fits the actual state of the building.

[0019] (2) Intelligence and pertinence of path planning enhancement According to the geometric characteristics of the building surface, independent cleaning areas are divided, and combined with the stain distribution density, vertical height, and area size, the priority score is calculated through a weighted formula to determine the cleaning order. It avoids the poor regional adaptability of the traditional uniform traversal strategy, prioritizes areas with high stain density and high operation difficulty, and realizes the rational allocation of resources such as unmanned aerial vehicle power and time.

[0020] For regions with different geometric characteristics (plane, curved surface, mixed structure), the optimal traversal direction is selected, the path segment spacing and length are dynamically adjusted according to the coverage range of the unmanned aerial vehicle cleaning head, the overlap rate of adjacent path segments is controlled at 10%-15%, and the Bezier curve is used to fit the path turning points to reduce the sharp turning action. It not only ensures full coverage of the cleaning area and avoids missed cleaning or repeated cleaning, but also reduces the energy consumption of the unmanned aerial vehicle, improves flight stability and operation efficiency.

[0021] (3) Obstacle detection and obstacle avoidance capability optimization Fusion of multi-modal data of RGB camera, depth camera and infrared thermal imaging module, real-time detection of obstacles through visual feature matching algorithm. Compared with traditional single sensor, it can effectively identify transparent, low contrast or special thermal characteristic obstacles, greatly reducing the miss detection rate and collision risk.

[0022] According to the relative relationship between the influence range of the obstacle and the length of the current path segment, the temporary jump or avoidance strategy is intelligently selected. For small obstacles, the temporary jump strategy is adopted to avoid low efficiency caused by excessive detour; for large obstacles, the shortest avoidance path is searched based on A* or RRT algorithm, and the safety distance is strictly set to ensure the safety and efficiency of obstacle avoidance, solving the problem of rigidity and dynamic response lag of traditional obstacle avoidance strategy.

[0023] (4) Overall operation efficiency improvement From precise modeling, intelligent path planning to real-time obstacle avoidance, the whole process is optimized to ensure that each area of the building surface can be effectively cleaned, especially in areas with high stain density and complex structures, improving the overall cleaning quality and uniformity.

[0024] Through reasonable resource allocation, efficient path planning and intelligent obstacle avoidance, the invalid flight and energy consumption of the unmanned aerial vehicle are reduced, the operation time and labor cost are reduced, and the endurance of the unmanned aerial vehicle is improved, which can support larger scale and higher frequency cleaning operation requirements.

[0025] Real-time obstacle detection and obstacle avoidance mechanism ensures the flight safety of the unmanned aerial vehicle in complex high-altitude environment, reduces the probability of equipment damage and accidents.

[0026] The application will be further described in detail below in combination with the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is the unmanned aerial vehicle high-altitude cleaning path precise planning method step diagram of the embodiment of the application; Figure 2 is the dynamic three-dimensional grid model generation flowchart of the embodiment of the application; Figure 3 is the cleaning path generation flowchart of the embodiment of the application; Figure 4 is the path obstacle real-time detection flowchart of the embodiment of the application; Figure 5 is the path re-planning flowchart of the embodiment of the application; Figure 6 is the unmanned aerial vehicle high-altitude cleaning path precise planning system module interaction diagram of the embodiment of the application.

[0028] Explanation of reference numerals: 21, unmanned aerial vehicle body; 211, flight control unit; 212, cleaning execution device; 22, multi-modal visual perception unit; 221, RGB camera; 222, depth camera; 223, infrared thermal imaging module; 23, initial model construction module; 24, dynamic model updating module; 25, path planning module; 26, obstacle detection and path re-planning module; 27, cleaning execution control module. DETAILED DESCRIPTION

[0029] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0030] It should be understood that the dimensions of the various elements shown in the drawings are shown for purposes of illustrating the application, and are in no way limiting.

[0031] The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the application or its application or uses.

[0032] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein. However, where appropriate, such techniques, methods, and apparatus should be considered as part of the description of the application.

[0033] Embodiment 1, refer to Figure 1 The application provides a UAV high-altitude cleaning path accurate planning method as shown in Figure 1 The application provides a UAV high-altitude cleaning path accurate planning method based on visual navigation, comprising the following steps: S101, preprocessing stage: constructing an initial three-dimensional grid model of the building to be cleaned, and dividing the cleaning area and setting the priority according to the initial three-dimensional grid model; S102, real-time sensing stage: the UAV collects visual data of the cleaning area through multi-modal visual sensors (RGB camera 221, depth camera 222, infrared thermal imaging module 223), including surface texture, depth information and stain distribution thermal map; S103, model dynamic updating stage: based on the real-time visual data, the initial three-dimensional grid model is locally corrected to generate a high-precision dynamic three-dimensional grid model; S104, initial path generation stage: using an adaptive grid traversal algorithm, according to the priority of the cleaning area and the coverage range of the UAV cleaning head, a cleaning path is generated; S105, dynamic path adjustment stage: real-time detection of obstacles in the path is performed through a visual feature matching algorithm, and when an obstacle is detected, the path is re-planned to avoid the obstacle until all areas are cleaned.

[0034] In the embodiment of the application, the preprocessing stage is the preliminary foundation preparation for path planning, the core is to establish the digital model of the building and clearly define the cleaning priority, and the subsequent efficiency is improved; the real-time sensing stage dynamically obtains accurate data of the cleaning area through multi-modal sensors, which provides a basis for model updating and path adjustment; the model dynamic updating stage corrects the initial model according to the real-time data, to ensure that the model is consistent with the actual state of the building; the initial path generation stage generates an initial cleaning path that adapts to the characteristics of the area, to ensure full coverage and high efficiency; the dynamic path adjustment stage detects obstacles in real time and adjusts the path to ensure safe cleaning.

[0035] BACKGROUND DESCRIPTION: In the high-altitude unmanned aerial vehicle cleaning scenario, accurate path planning depends on the digital perception of the geometric structure and texture features of the building surface, and the construction of the initial three-dimensional grid model is the core basis of this process. The traditional three-dimensional modeling method of buildings (such as manual measurement, static CAD model) has the pain points of low efficiency, poor adaptability (unable to cope with dynamic changes of building surface or complex protruding structures), high risk coefficient of high-altitude operation, etc. Based on this: In the above step S101, when constructing the initial three-dimensional grid model of the building to be cleaned, it includes: The RGB camera 221 carried by the pre-flight unmanned aerial vehicle collects color images of the building surface, and the depth camera 222 collects distance information, and a rough three-dimensional point cloud model is generated by combining a simultaneous localization and mapping (SLAM) algorithm, and then converted into an initial three-dimensional grid model through grid simplification and texture mapping.

[0036] In the embodiments of the present application, it needs to be further explained that the building surface contains complex geometric shapes (plane, curved surface, corner) and texture features (material difference, stain distribution), and a single sensor cannot fully restore it. The RGB camera 221 captures color texture information, and the depth camera 222 obtains spatial distance data, and the combination of the two can complementarily generate original data with visual details and three-dimensional structure, providing complete input for subsequent model construction.

[0037] In the high-altitude environment, GPS signals are easily blocked or interfered, and cannot rely on external positioning. The SLAM algorithm can complete the self-positioning and building surface map construction in real time during the movement of the unmanned aerial vehicle, and autonomously generate a three-dimensional point cloud model containing a large number of discrete points, without relying on a pre-set map or manual marking, to realize an efficient and autonomous modeling process.

[0038] Discrete point cloud data is large in amount and high in computational complexity, and direct use in path planning will result in insufficient real-time performance. Through grid simplification to reduce redundant data (key structure boundaries are retained), and texture mapping to attach RGB texture to the grid surface, the generated initial three-dimensional grid model is both lightweight (suitable for real-time calculation) and can accurately reflect the geometric boundaries and material features of the building, providing a reliable digital reference for subsequent cleaning area division, priority setting and dynamic path adjustment, and is a key bridge connecting physical buildings and intelligent path planning.

[0039] The design of this step effectively solves the limitations of traditional modeling methods and lays a digital foundation for accurate path planning of high-altitude unmanned aerial vehicle cleaning.

[0040] BACKGROUND DESCRIPTION: In the high-altitude cleaning scenario, the building surface often contains complex geometric structures (mixing of plane, curved surface, corner) and dynamic stain distribution, and the traditional path planning has two limitations: Non-targeted regional processing: Traditional methods mostly use uniform traversal strategies (such as global snake shape), which cannot adapt to regions with different geometric characteristics (such as curved walls, which are easily missed by plane traversal); Ambiguous priority: Artificially setting cleaning order is easily influenced by subjective factors, often ignoring the priority processing of high-stain and high-difficulty regions (such as high-altitude vertical surfaces), leading to resource waste (such as low-stain regions being cleaned first, and high-difficulty regions not being covered due to insufficient energy consumption). Based on this: In the above step S101, when the cleaning regions are divided and the priority is set according to the initial three-dimensional mesh model, it includes: Cleaning region division: Based on the surface geometric characteristics (such as plane attributes, curved surface attributes, and mixed attributes) of the initial three-dimensional mesh model, the building surface is divided into multiple independent cleaning regions; Priority setting: Using pre-flight infrared thermal imaging data or color image texture analysis, the average stain distribution density (such as the proportion of dark stains, the proportion of abnormal thermal radiation regions) of each region is preliminarily evaluated, and combined with the vertical height and area size of the region, a weighted formula is used to calculate the priority score (for example: priority score = stain distribution density average × 0.5 + vertical height × 0.3 + area size × 0.2); According to the priority score from high to low, the cleaning priority order of each cleaning region is determined.

[0041] In the embodiments of the present application, it needs to be further explained that (1) the principle of cleaning region division The geometric differences of the building surface directly determine the adaptability of the cleaning path: flat areas are suitable for snake traversal, curved areas (such as arc-shaped walls) are suitable for ring traversal, and mixed geometric characteristic areas need segmented adaptive strategies. If not segmented, uniform path planning may lead to the following problems: Missed cleaning or repeated cleaning: forced use of flat snake paths in curved areas will result in cleaning blind spots due to insufficient adhesion; Energy waste: the device frequently switches between different geometric regions, increasing unnecessary energy consumption; Precision decline: unable to adjust the cleaning head angle and coverage range for complex structures (such as corners, cylinders).

[0042] Therefore, based on the surface geometric characteristics (plane / curve / mixed attributes) of the initial three-dimensional mesh model, independent regions are segmented, which is the premise of subsequent adaptive path planning: through regional processing, ensure that the path strategy of each region is highly matched with its own form, improve cleaning accuracy and efficiency.

[0043] (2) The principle of priority setting The core logic of priority setting is to optimize resource allocation (drone power, operation time), to ensure that high-value and high-demand areas are processed in priority, and the specific basis is shown in Table 1: Table 1

[0044] The weight distribution of the weighted formula (stain 0.5 > height 0.3 > area 0.2) is based on the priority of the actual cleaning demand: the degree of dirtiness is the core consideration (directly affecting the cleaning effect), followed by the safety risk (high-altitude operation priority), and finally the efficiency optimization (reducing path repetition). Through score ranking, resources are preferentially directed to the areas that need them most, avoiding the "inefficient cleaning" caused by traditional non-priority planning (such as cleaning clean areas first and dirty areas running out of power later).

[0045] Background description: In the field of unmanned aerial cleaning path planning, traditional technologies have core pain points such as single perception method, missing data dimensions, and poor scene adaptability, which seriously restrict the accuracy and efficiency of cleaning operations. Based on this: In the above step S102, the visual data of the cleaning area is collected, including surface texture, depth information and stain distribution heat map, which includes: Surface texture data collection: The RGB camera 221 collects color images of the cleaning area at a preset frame rate, and performs texture feature extraction (such as gray level co-occurrence matrix, local binary pattern algorithm, etc.) on the images to generate a surface texture feature map containing color distribution and texture complexity; Depth information collection: The depth camera 222 acquires distance data of each point in the cleaning area relative to the drone in real time, and combines the real-time pose information of the drone (output by the simultaneous localization and mapping (SLAM) algorithm) to calculate the three-dimensional spatial depth coordinates of each point in the cleaning area; Stain distribution heat map generation: The infrared thermal imaging module 223 collects thermal radiation intensity data of the cleaning area, and maps the thermal radiation value to a heat map with different color gradients (thermal radiation abnormal area corresponds to high stain probability); At the same time, combined with the proportion analysis result of dark stain area in the color image collected by the RGB camera 221, the heat map is corrected to obtain a stain distribution heat map that accurately reflects the stain distribution.

[0046] In the embodiments of the present application, it needs to be further explained that step S102 synchronously collects surface texture, depth information and stain distribution heat map through multi-modal visual sensor fusion technology, to provide accurate data support for subsequent dynamic model correction and intelligent path planning, and the specific effects are as follows: (1) Surface texture collection: restore the real details of the building surface The texture characteristics of the building surface, such as material differences and wear levels, directly affect the adaptability of the cleaning path: Coarse texture areas, such as stone walls, require increased cleaning head pressure or slower flight speed; Smooth texture areas, such as glass curtain walls, require reduced overlap rates to avoid water waste.

[0047] Color images are captured by an RGB camera and texture features (gray level co-occurrence matrix, local binary pattern) are extracted to generate a surface texture feature map, providing a basis for subsequent dynamic adjustment of path parameters.

[0048] (2) Depth information collection: building precise three-dimensional spatial coordinates In high-altitude environments, GPS signals are easily blocked, and the UAV needs to rely on autonomous perception to achieve positioning and modeling: Depth cameras obtain real-time distance data between each point in the cleaning area and the UAV; Combined with the real-time pose information of the UAV output by the SLAM algorithm, the distance data is converted into three-dimensional spatial depth coordinates, providing a geometric reference for local correction of the initial three-dimensional mesh model.

[0049] (3) Stain distribution heat map: achieving fine cleaning priority management Traditional stain detection cannot accurately quantify the degree and distribution range of dirt, leading to unreasonable allocation of cleaning resources: The infrared thermal imaging module identifies high-probability stain areas based on differences in thermal radiation intensity (stains have obvious differences in thermal radiation values compared to clean areas due to characteristics such as dust accumulation and water absorption); Combined with the dark stain proportion analysis results of the RGB image, the heat map is corrected to effectively exclude interference factors such as shadows and fading, generating a heat map that accurately reflects stain distribution, providing data support for subsequent priority cleaning of high-stain density areas.

[0050] This step breaks through the information limitations of traditional single sensors through the fusion of multi-modal data, achieving comprehensive and accurate perception of the geometric structure, texture details, and stain distribution of building surfaces, providing a core data foundation for subsequent dynamic model updating, adaptive path planning, and intelligent obstacle avoidance. It is a key link to improve the accuracy, efficiency, and safety of high-altitude UAV cleaning operations.

[0051] Background description: In high-altitude UAV cleaning operations, accurate path planning needs to meet three major goals: full coverage without missing cleaning, high efficiency and energy saving, and safe obstacle avoidance. Traditional methods rely on static CAD models or manual experience, which cannot handle real-time changes in building surfaces (such as temporary obstacles and uneven stain distribution) and complex geometric features (such as curved walls and protruding structures), resulting in low cleaning efficiency, high missing cleaning rate, or high collision risk. Based on this: Reference Figure 2a dynamic three-dimensional grid model generation flowchart, in the step S103, when the high-precision dynamic three-dimensional grid model is generated, comprises: Data coordinate system alignment: the real-time collected surface texture feature map, three-dimensional space depth coordinate and stain distribution heat map are converted to the unified coordinate system of the initial three-dimensional grid model through the real-time pose information of the unmanned aerial vehicle (output by the simultaneous localization and mapping (SLAM) algorithm), so that the data space consistency is ensured; Local difference detection: the deviation of the geometric parameters (such as grid vertex position, point cloud density) of the corresponding area of the initial three-dimensional grid model and the three-dimensional space depth coordinate data, and the detail difference between the initial texture mapping and the surface texture feature map are compared, and the area with a deviation exceeding a preset threshold is selected as a to-be-revised area; Geometric structure correction: for the to-be-revised area, the three-dimensional space depth coordinate data is used to adjust the grid vertex position, and the local grid topology structure (such as subdividing the grid or adjusting the connection relationship of the triangular face) is optimized to match the actual geometric shape of the building surface; Texture and stain information updating: the surface texture feature map is used to replace the initial texture mapping of the to-be-revised area to retain the real surface details; meanwhile, the high-stain-density area in the stain distribution heat map is marked as a key area of the dynamic three-dimensional grid model, and the accuracy of the geometric and texture details thereof is strengthened; The setting of the preset threshold is based on: the geometric deviation threshold does not exceed 5% of the coverage range of the unmanned aerial vehicle cleaning head, and the texture difference threshold is less than 0.8 (or the matching rate of the local binary pattern is less than 70%) based on the similarity of the gray level co-occurrence matrix. When the local grid topology structure is optimized, the grid integrity of the building structure boundary (such as the window edge and the wall corner) is preferentially retained to ensure that the corrected model can still accurately reflect the building functional area division.

[0052] In the embodiments of the present application, it needs to be further explained that the generation of the dynamic three-dimensional grid model is a core link connecting the initial static model and the real-time cleaning scene, and through real-time data fusion correction, the model is ensured to be highly consistent with the actual state of the building, thereby providing an accurate digital reference for subsequent path planning.

[0053] (1) Data coordinate system alignment Principle: the data collected by the multi-modal sensors (RGB camera 221, depth camera 222, infrared module) is in the respective coordinate systems (such as the camera itself coordinate system and the unmanned aerial vehicle body coordinate system) by default, and the initial three-dimensional grid model is constructed based on the world coordinate system. If the coordinate systems are not unified, the real-time data cannot be corresponded to the initial model, resulting in correction deviation.

[0054] The real-time pose information (position: x / y / z coordinates; attitude: pitch / roll / yaw angle) output by the UAV SLAM algorithm is used to convert the real-time collected surface texture feature map, three-dimensional space depth coordinates, and stain distribution heat map to the unified world coordinate system of the initial three-dimensional mesh model through coordinate transformation (rotation + translation matrix).

[0055] Ensure that all real-time data correspond to the spatial position of the initial model, laying the foundation for subsequent difference detection and correction.

[0056] (2) Local difference detection Principle: The initial three-dimensional mesh model is generated from pre-flight data, which may have deviations from the actual state of the building (such as minor surface deformation, insufficient initial modeling accuracy). By comparing the differences between real-time data and the initial model, areas that need to be corrected are selected, avoiding the computational redundancy of global model updating.

[0057] Geometric difference detection: Compare the deviations of the grid vertex positions, point cloud density of the corresponding area of the initial three-dimensional mesh model, and real-time three-dimensional space depth coordinates. If the deviation exceeds "5% of the coverage range of the UAV cleaning head" (threshold setting basis: ensure that the accuracy of the corrected area is sufficient to support the accuracy of the cleaning path), mark it as a geometric area to be corrected.

[0058] Texture difference detection: Compare the detailed differences between the initial texture mapping and the real-time surface texture feature map. Use the gray level co-occurrence matrix (GLCM) similarity (less than 0.8) or local binary pattern (LBP) matching rate (less than 70%) as the threshold to select the texture area to be corrected.

[0059] Finally, areas with geometric or texture deviations exceeding the threshold are determined as areas to be corrected.

[0060] (3) Geometric structure correction Principle: For the areas to be corrected, use real-time high-precision depth data to adjust the geometric structure of the initial model to fit the actual shape of the building; while preserving the key structure boundaries to ensure the accuracy of the functional area division of the model.

[0061] Vertex position adjustment: Use real-time three-dimensional space depth coordinate data to correct the x / y / z coordinates of the grid vertices in the area to be corrected, so that the grid is consistent with the actual geometric shape of the building surface.

[0062] Topology optimization: For curved or complex structure areas, use subdivision mesh (increase the number of triangular patches) or adjust the connection relationship of triangular patches to improve the model's ability to express complex shapes; for flat areas, simplify the adjustment to balance accuracy and computational efficiency.

[0063] Boundary Integrity Preservation: Prioritize preserving the grid integrity of architectural structure boundaries (e.g., window edges, wall corners) to avoid disrupting the division of functional areas (e.g., the boundaries between windows and walls) during the correction process.

[0064] (4) Texture and Stain Information Update Principle: Real-time texture data is more realistic than the pre-flight texture of the initial model, replacing it can improve the visual detail accuracy of the model; mark high-stain areas as priorities to strengthen their geometric and texture details, providing accurate dirty distribution reference for subsequent path planning.

[0065] Texture Replacement: Replace the initial texture mapping of the area to be corrected with the real-time collected surface texture feature map (including color distribution, texture complexity), preserving the real material details of the building surface (such as tile gaps, paint peeling, etc.).

[0066] Stain Area Marking: Mark high-stain density areas (such as areas with high abnormal heat radiation and high dark stain proportion) in the real-time generated stain distribution heat map as the focus areas of the dynamic model, and perform additional optimization (such as subdividing the grid and improving texture resolution) on the grid vertex position and texture details of these areas to ensure that the subsequent cleaning path can accurately cover these areas.

[0067] Through the above steps, the dynamic three-dimensional grid model realizes the iterative update of "initial model → real-time data correction → high-precision dynamic model", providing a reliable digital foundation for subsequent adaptive path planning and obstacle avoidance.

[0068] Background Description: Traditional high-altitude cleaning path planning often uses random or globally uniform regional processing order, lacking clear priority division logic. This approach can lead to unreasonable allocation of resources (drone power, operation time); secondly, traditional path planning often uses fixed traversal directions (such as global snake traversal), which cannot adapt to the complex and diverse geometric features of building surfaces; in addition, traditional path planning does not fully consider the actual coverage range of the drone cleaning head, and the path segment spacing and length are fixed, which can lead to two extreme problems: too large spacing: leading to missed cleaning areas, requiring secondary cleaning; too small spacing: high overlap rate (more than 15%), increasing invalid operation time and energy consumption; finally, traditional path planning often uses right angles or sharp turns for turning points, resulting in two major problems: Increased energy consumption: sharp turns require the drone to quickly adjust its attitude, significantly increasing energy consumption; Decreased stability: in complex high-altitude wind fields, frequent sharp turns can lead to loss of control and increase safety risks.

[0069] In addition, frequent direction changes can disrupt the continuous cleaning process, reducing operation efficiency. Based on this: Reference Figure 3The cleaning path generation flowchart in step S104 includes: Cleaning area allocation: according to the cleaning priority order of each cleaning area, the current cleaning area is allocated; Adaptive selection of traversal direction: according to the surface geometric characteristics (such as planar attribute, curved surface attribute and mixed attribute) of the cleaning area, the optimal traversal direction is determined: Planar area: the direction consistent with the long axis of the area is selected for serpentine traversal; Curved surface area (such as arc wall surface, cylindrical structure): ring traversal along the normal direction or circumferential direction of the curved surface is selected; Mixed geometric feature area: segmented traversal is adopted to adapt to the geometric shape of different sub-areas; Adjustment of coverage range: according to the coverage range (such as diameter D) of the unmanned aerial vehicle cleaning head, the spacing and length of the path segment are adjusted to control the overlap rate of adjacent path segments at 10%-15%, ensuring no missed cleaning; Path smoothing optimization: Bezier curve fitting is adopted for path turning points to reduce sharp turning actions; paths in the same direction are preferentially selected to reduce direction switching frequency and optimize energy consumption.

[0070] In the embodiments of the present application, (1) the cleaning area allocation mechanism Principle: based on the cleaning priority order in the preprocessing stage, high-value areas are preferentially allocated to optimize resource (unmanned aerial vehicle power, operation time) allocation efficiency.

[0071] Specific mechanism: Priority source: the preprocessing stage calculates the area priority score (stain distribution density mean × 0.5 + vertical height × 0.3 + area size × 0.2) by a weighting formula, and the higher the score, the higher the priority.

[0072] Allocation logic: high-priority areas (such as areas with severe dirt, high vertical height and large area) are preferentially processed to avoid incomplete cleaning due to insufficient power or repeated paths, and to maximize the value of high-altitude operation of the unmanned aerial vehicle (such as preferentially cleaning high-altitude dirty areas that are difficult for humans to reach).

[0073] (2) Adaptive selection mechanism of traversal direction Principle: according to the surface geometric characteristics of the area, the traversal direction highly matched with the shape is selected to ensure coverage accuracy and efficiency.

[0074] Specific mechanism: Planar area: the serpentine traversal direction consistent with the long axis of the area is selected.

[0075] Logic: Serpentine traversal path is continuous and has less direction switching, which can maximize the use of the long axis space of the area, reduce invalid flight, and improve coverage uniformity.

[0076] Curved surface area (such as arc wall surface, cylindrical structure): Choose ring traversal along the normal direction of the curved surface or the circumferential direction.

[0077] Logic: Fit the curved surface contour to avoid missing cleaning caused by plane traversal (such as using ring traversal to completely cover the circumference of the cylindrical surface), and ensure that the path is consistent with the curved surface shape.

[0078] Mixed geometric feature area: Use segmented traversal to adapt to the geometric shape of different sub-areas.

[0079] Logic: Mixed areas contain various shapes such as planes and curved surfaces. Segmented processing can choose the optimal direction for each sub-area to avoid precision decline caused by uniform traversal.

[0080] (3) Coverage range adjustment mechanism Principle: Adjust the path parameters according to the actual coverage range of the unmanned aerial vehicle cleaning head to balance coverage completeness and operation efficiency.

[0081] Specific mechanism: Path parameter adjustment: Adjust the distance and length of the path segment according to the coverage range of the cleaning head (such as diameter D).

[0082] Overlap rate control: Keep the overlap rate of adjacent path segments at 10%-15%.

[0083] Logic: Low overlap rate (<10%): Easy to miss cleaning areas and need secondary cleaning; High overlap rate (>15%): Increases invalid operation time and energy consumption; 10%-15% is the optimal interval that balances coverage completeness and efficiency, ensuring no missed cleaning and minimal resource waste.

[0084] (4) Path smoothing optimization mechanism Principle: Reduce the energy consumption and risk of unmanned aerial vehicle attitude adjustment, improve flight stability and operation efficiency.

[0085] Specific mechanism: Turning point smoothing: Use Bezier curve to fit the turning points of the path.

[0086] Logic: Bezier curve can convert right-angle or sharp turning points into smooth curves, reducing the sharp movements of the unmanned aerial vehicle, reducing the energy consumption of attitude adjustment, and improving the stability of high-altitude flight (avoiding the risk of loss of control under wind field interference).

[0087] Direction switching optimization: prioritize paths in the same direction to reduce the frequency of direction switching.

[0088] Logic: Direction switching requires the drone to adjust its flight attitude, which consumes a significant amount of power. Continuous movement in the same direction reduces the number of switches, optimizes energy consumption, and improves the continuity of cleaning operations.

[0089] The above four aspects work together to form a cleaning path generation logic that adapts to building features, is efficient and low in energy consumption, providing core support for the precision and safety of high-altitude drone cleaning.

[0090] Background description: In the scenario of high-altitude drone cleaning operations, real-time obstacle detection is a key step to ensure flight safety and cleaning efficiency. Traditional high-altitude cleaning path planning relies on static CAD models or manual experience, which has two major pain points: first, it cannot cope with dynamic changes on the building surface (such as temporary scaffolding, billboards, air conditioning units, and other sudden obstacles); second, the sensing ability of a single sensor is limited (such as using only an RGB camera 221, which is prone to miss transparent or low-contrast obstacles, and using only a depth camera 222, which cannot distinguish between wall texture and real obstacles), resulting in high collision risk or cleaning interruption. At the same time, GPS signals in high-altitude environments are easily blocked, and external positioning systems cannot be relied upon for obstacle avoidance, so obstacle detection must be done through autonomous sensing by the drone. Based on this: Referring to Figure 4 the path obstacle real-time detection flowchart, in step S105, when detecting obstacles in the path in real time through a visual feature matching algorithm, it includes: Real-time multi-modal feature extraction: Extract local feature points from the real-time frames of the drone's RGB camera 221 (using ORB or SIFT algorithms), simultaneously obtain the three-dimensional spatial depth coordinates corresponding to the local feature points from the depth camera 222, and obtain the thermal radiation intensity values corresponding to the local feature points from the infrared thermal imaging module 223; Feature library matching and deviation calculation: Match the extracted local feature points with the pre-stored feature library (including geometric position, depth reference, and thermal radiation reference value) of the corresponding area in the dynamic three-dimensional grid model, and calculate the position deviation ΔP, depth deviation ΔD, and thermal radiation deviation ΔT of the feature points; Obstacle determination threshold setting: Set the obstacle determination threshold, where the position deviation ΔP threshold does not exceed 10% of the coverage range of the drone's cleaning head, the depth deviation ΔD threshold does not exceed 5 cm, and the thermal radiation deviation ΔT threshold does not exceed 20% of the normal reference value of the area; when any deviation exceeds the corresponding threshold, mark the area where the feature point is located as a suspected obstacle area; Obstacle area verification and positioning: For suspected obstacle areas, through continuous 3-frame or more visual data tracking verification, after confirming the existence of obstacles, the three-dimensional spatial depth coordinates, bounding rectangle size and contour shape of the obstacles are calculated using depth camera 222 data and sent to the path planning unit to provide accurate input for dynamic path adjustment.

[0091] Further explained in the embodiments of the present application is that the embodiments realize autonomous and accurate detection of obstacles in high-altitude cleaning scenarios through the logic of multi-modal perception fusion + dynamic model comparison + threshold determination + multi-frame verification, which solves the pain points of traditional single sensor missed detection and static model unable to cope with dynamic changes, and provides reliable input for path re-planning.

[0092] (1) Real-time multi-modal feature extraction: multi-dimensional data complement each other to eliminate perception blind spots Principle: In a high-altitude environment, a single sensor cannot comprehensively identify obstacles (e.g., an RGB camera 221 is prone to miss transparent glass / low-contrast objects, a depth camera 222 is difficult to distinguish wall texture from real obstacles, and an infrared camera can capture thermal radiation anomalies). Therefore, through the multi-modal combination of RGB camera 221 + depth camera 222 + infrared thermal imaging module 223, features are extracted from three dimensions: RGB camera 221: extract local feature points (such as corners, texture discontinuities) using ORB / SIFT algorithm to capture object visual contours; Depth camera 222: synchronously obtains three-dimensional spatial coordinates of feature points to provide distance information; Infrared module: obtains thermal radiation intensity values of feature points to identify thermal anomaly areas (such as the thermal difference between metal obstacles and walls).

[0093] The three data complement each other, covering visual, spatial, and thermal dimensions, reducing missed detection and misjudgment.

[0094] (2) Feature library matching and deviation calculation: compare dynamic model benchmarks to identify abnormal changes Principle: The dynamic three-dimensional grid model pre-stores the geometric position, depth benchmark, and thermal radiation benchmark (i.e., the "normal state" feature library) of the corresponding area. Match the real-time extracted feature points with the library and calculate three key deviations: Position deviation ΔP: the difference between the three-dimensional coordinates of the real-time feature points and the model benchmark coordinates; Depth deviation ΔD: the difference between the real-time depth value and the model benchmark depth; Thermal radiation deviation ΔT: the difference between the real-time thermal radiation value and the model benchmark value.

[0095] The existence of obstacles will cause these deviations to significantly exceed the normal range (e.g., a sudden increase in depth deviation corresponds to a newly added object on the wall).

[0096] (3) Obstacle determination threshold setting: Quantize abnormal boundaries and screen suspected areas Principle: Three thresholds are preset. If any deviation exceeds the threshold, it is marked as a suspected obstacle area. The threshold setting is based on actual cleaning needs: Position deviation ΔP≤10% of cleaning head coverage: Ensure that the deviation is significant enough (not a minor error) to avoid false positives due to small position shifts, while ensuring that detected obstacles do not affect the cleaning path or cause collisions; Depth deviation ΔD≤5cm: Small deviations may be sensor noise, and deviations exceeding 5cm indicate the presence of a physical obstruction (such as a protruding air conditioner outdoor unit); Thermal radiation deviation ΔT≤20%: The difference in thermal radiation characteristics between obstacles (such as metal and plastic) and walls usually exceeds this range, allowing for the identification of temperature anomalies (such as heat-emitting equipment).

[0097] Any deviation exceeding the threshold is marked as a suspected obstacle, covering various types of obstacles such as physical obstructions and temperature anomalies, improving overall detection.

[0098] (4) Obstacle area verification and positioning: Multi-frame tracking confirmation, output accurate information Principle: Single-frame data is susceptible to noise interference (such as light and shadow changes), so suspected areas are verified through continuous tracking of 3 or more frames: If the same deviation is detected in multiple consecutive frames, it is confirmed that an obstacle exists; Use depth camera 222 data to calculate the three-dimensional spatial coordinates, bounding rectangle size, and contour shape of the obstacle, and synchronize to the path planning unit.

[0099] Multi-frame verification eliminates false positives, and three-dimensional information provides accurate input on "where the obstacle is and how big it is" for subsequent path re-planning, ensuring the safety and adaptability of the obstacle avoidance path.

[0100] In summary, this process achieves autonomous detection of obstacles in high-altitude environments through a closed loop of "multi-modal perception → dynamic model comparison → quantitative determination → accurate verification", providing a core guarantee for the safe cleaning of drones.

[0101] Background Description: In high-altitude unmanned aerial cleaning operations, real-time obstacle avoidance and path re-planning are the core links to ensure safety and efficiency. There are three key pain points in traditional high-altitude cleaning path planning: first, the obstacle avoidance strategy is single, regardless of the size or influence range of the obstacle, a unified detour method is adopted, which leads to path redundancy in small obstacle scenarios (such as small objects protruding locally on the wall, excessive detour will increase the invalid flight time and energy consumption), and incomplete obstacle avoidance in large obstacle scenarios (such as not fully considering the safety distance around the obstacle, which is prone to collision risk); second, the safety distance setting is unreasonable, the traditional method often ignores the adaptability of the obstacle and the cleaning head of the unmanned aerial vehicle, the safety distance is too small (which cannot avoid collision) or too large (which wastes path resources); third, the dynamic adaptability is poor, relying on static building models, which cannot cope with temporary obstacles (such as suddenly added scaffolds, billboards, air conditioner outdoor units, etc.) in high-altitude scenarios, leading to cleaning interruption or missed cleaning. At the same time, GPS signals are easily blocked in high-altitude environments, and the unmanned aerial vehicle needs to rely on autonomous perception and dynamic models for path adjustment, further requiring the obstacle avoidance strategy to have precision and real-time performance. Based on this: Referring to Figure 5 the path re-planning flowchart, when the obstacle is detected in step S105, the path re-planning is performed, which includes: Obstacle influence range definition: using the obtained three-dimensional space depth coordinates of the obstacle, the size of the circumscribed rectangle, and the contour shape, combined with the geometric boundaries of the corresponding region in the dynamic three-dimensional grid model, the obstacle influence range (including the obstacle body and the surrounding safety distance region, the safety distance is not less than 20% of the coverage range of the unmanned aerial vehicle cleaning head) is determined; Adaptive selection of re-planning strategy: according to the relative position relationship between the obstacle influence range and the current cleaning path, the optimal strategy is selected: If the obstacle influence range is less than 30% of the length of the current path segment: adopt the temporary jump strategy, adjust the position of the end point of the current path segment, jump over the obstacle influence range and then restore the original traversal direction; If the obstacle influence range is greater than or equal to 30% of the length of the current path segment: adopt the detour avoidance strategy, based on A* algorithm or Rapidly-exploring Random Tree (RRT) algorithm, search for the shortest detour path from the current position to the target position on the other side of the obstacle in the dynamic three-dimensional grid model, and restore the original traversal direction after completing the detour.

[0102] In the embodiments of the present application, it needs to be further explained that the present embodiment solves the limitations of traditional obstacle avoidance through two core links of obstacle influence range definition and adaptive strategy selection: (1) Obstacle influence range definition: combined with the geometric boundary of the dynamic three-dimensional grid model, the obstacle body and the surrounding safety distance (not less than 20% of the cleaning head coverage range) are included in the influence range, which not only ensures collision safety (avoids scratching between the unmanned aerial vehicle and the obstacle or protruding structure of the building), but also avoids the efficiency loss caused by excessive expansion of the obstacle avoidance area; (2) Adaptive strategy selection: according to the ratio of the obstacle influence range to the current path segment length, dynamically switch between temporary jump or avoidance strategy: small influence range (<30% path segment length) uses temporary jump strategy (resumes the original traversal direction after jumping over the obstacle, reduces redundant path), large influence range (≥30%) uses A* or RRT algorithm to avoid (ensures to completely avoid the obstacle, while searching for the shortest path), realizes the balance of safety and efficiency.

[0103] The above design logic aims to enable the unmanned aerial vehicle to accurately avoid temporary obstacles in the high-altitude dynamic environment, while maintaining the continuity and efficiency of the cleaning path to the greatest extent, solving the problem of resource waste and safety hazards caused by the "one-size-fits-all" of traditional obstacle avoidance strategies.

[0104] In one possible embodiment, when searching for the shortest avoidance path from the current position to the target position on the other side of the obstacle in the dynamic three-dimensional grid model, it includes: Set search constraint boundary: based on the three-dimensional space depth coordinates, the circumscribed rectangle size and the contour shape of the obstacle in the dynamic three-dimensional grid model, combined with the retained building structure boundary (such as window edge, wall corner) in the model, it is clear that the path cannot pass through the obstacle body and the building structure boundary, and the minimum safety distance between the path and the obstacle is not less than 20% of the cleaning head coverage range of the unmanned aerial vehicle; Construct a multi-dimensional cost function, including: path total length (weight coefficient 0.4), safety distance between path and obstacle (weight coefficient 0.3), curvature change rate of path turning point (weight coefficient 0.2, using Bezier curve fitting to evaluate smoothness) and overlap degree of path coverage area and high-stain density area in stain distribution heat map (weight coefficient 0.1); Perform constraint path search: use A* algorithm or rapid extended random tree (RRT) algorithm to generate an initial avoidance path based on the multi-dimensional cost function in the search space corresponding to the dynamic three-dimensional grid model, with the current position as the starting point and the target position on the other side of the obstacle as the ending point; Path optimization and adaptation: smooth the initial avoidance path (smooth the turning points by Bezier curve fitting), and adjust the path segment spacing to ensure that the overlap rate of adjacent path segments is maintained at 10%-15%; Path verification and update: verify whether the optimized path meets the constraint conditions and cleaning coverage requirements, if it meets the requirements, update it to the path planning unit, and mark it in the dynamic three-dimensional grid model at the same time.

[0105] Further explained in the embodiments of the present application is that the embodiments realize safe, efficient and high-quality path planning through the logical chain of "constraint boundary definition → multi-objective cost trade-off → intelligent path search → practical optimization → verification closed loop", and the principles of each link are as follows: (1) Search constraint boundary: prerequisite for safety and feasibility Principle: Strictly define the "safety framework" for path search to avoid collision risks.

[0106] Prohibit crossing area: Combine the three-dimensional coordinates, size and building structure boundary (such as window edge, wall corner) of the obstacle to clearly define the space range that the path cannot enter, which prevents the unmanned aerial vehicle from colliding with the obstacle body and also avoids scratching the inherent structure of the building (such as the edge of the glass curtain wall).

[0107] Minimum safety distance: Set as "20% of the cleaning head coverage range", which is based on the operation characteristics of the unmanned aerial vehicle: the cleaning head coverage range is a known operating unit, and the distance of 20% can ensure that the unmanned aerial vehicle maintains a safe buffer from the obstacle (to avoid collision caused by air flow disturbance) and does not excessively expand the avoidance range (to reduce invalid flight).

[0108] (2) Multi-dimensional cost function: decision basis for comprehensive optimal path Principle: Through the weighted fusion of multiple targets, guide the algorithm to select the path of "safety, efficiency, energy saving and high quality", rather than a single "shortest path".

[0109] Total path length (weight 0.4): highest priority, because the shortest path directly reduces energy consumption and operation time, which meets the endurance limit of high-altitude operation of the unmanned aerial vehicle.

[0110] Safety distance (weight 0.3): secondary priority, which ensures that the distance between the path and the obstacle is always greater than the safety threshold, and is the core guarantee for high-altitude operation safety.

[0111] Curvature change rate (weight 0.2): evaluate the smoothness of the path through the Bezier curve, the smaller the curvature (the fewer the sharp turns), the more stable the unmanned aerial vehicle flies (reducing the risk of loss of control under wind field interference), and the lower the energy consumption.

[0112] High-stain area overlap degree (weight 0.1): consider cleaning quality, try to cover high-stain areas when detouring to avoid missing key dirty parts due to detouring.

[0113] Logic: The weighted mechanism balances the "efficiency (length), safety (distance), stability (curvature) and quality (overlap degree)" four targets, so that the path generated by the algorithm is more in line with the actual operation requirements.

[0114] (3) Constraint path search: application of intelligent algorithm Principle: Combine constraint boundaries with cost functions to quickly find feasible and optimal initial paths using mature path planning algorithms.

[0115] A* Algorithm: Suitable for static and known environments (such as fixed building structures), can efficiently find the optimal path with the smallest cost.

[0116] RRT Algorithm: Suitable for dynamic and complex environments (such as temporary obstacles), can quickly explore the space to find a feasible path.

[0117] Logic: Within the constraint boundary, the algorithm takes the "current position" as the starting point and the "target position on the other side of the obstacle" as the ending point, calculates the cost of each step based on the cost function, and selects the path with the smallest cost as the initial avoidance path.

[0118] (4) Path optimization adaptation: from "feasible" to "practical" upgrade Principle: Perform engineering optimization on the initial path to make it meet the flight characteristics and cleaning requirements of the UAV.

[0119] Smooth processing: Use Bezier curves to fit turning points, convert right angles / rapid turns into smooth curves, reduce energy consumption and stability risks of UAV attitude adjustment (rapid turns are prone to loss of control in high-altitude wind fields).

[0120] Coverage adjustment: Maintain a 10%-15% overlap rate for path segments, which is a key indicator for cleaning coverage, avoiding both large gaps leading to missed cleaning and small gaps causing repeated cleaning (wasting time and water resources).

[0121] (5) Path verification and update: reliability closed-loop protection Principle: Ensure the correctness of the path through verification, and update the model to support subsequent planning.

[0122] Constraint verification: Check if the path meets the safety distance, building boundary and other constraint conditions to avoid "theoretically feasible but actually dangerous" paths.

[0123] Coverage verification: Confirm that the overlap rate meets the standard to ensure cleaning quality.

[0124] Model update: Mark the verified path to the dynamic three-dimensional grid model, and subsequent path planning can directly reference the position of this obstacle to avoid repeated detection and avoidance errors.

[0125] The core principle of this embodiment is "multi-objective optimization under constraints": taking the safety boundary as the premise, balancing multiple operation objectives through the cost function, generating the initial path using intelligent algorithms, and finally outputting a safe and efficient avoidance path through engineering optimization and verification. The whole process closely revolves around the scene pain points of high-altitude UAV cleaning (high safety risk, limited endurance, and strict cleaning quality requirements), and realizes the precise landing of "from digital model to actual operation".

[0126] Embodiment 2, see Figure 6 The present application provides a UAV high-altitude cleaning path precise planning system module interaction diagram as shown in Figure 6 The present application provides a UAV high-altitude cleaning path precise planning system based on visual navigation, which comprises a UAV body 21, a multi-modal visual perception unit 22, an initial model construction module 23, a dynamic model updating module 24, a path planning module 25, an obstacle detection and path re-planning module 26, and a cleaning execution control module 27.

[0127] The UAV body 21 is equipped with a flight control unit 211 and a cleaning execution device 212; The multi-modal visual perception unit 22 includes an RGB camera 221, a depth camera 222, and an infrared thermal imaging module 223, which are used to collect the surface texture, depth information, and stain distribution thermal map of the cleaning area; The initial model construction module 23 is used to generate an initial three-dimensional grid model by combining the image and distance information collected by the UAV pre-flight and the simultaneous localization and mapping (SLAM) algorithm, and to divide the cleaning area and set the priority; The dynamic model updating module 24 locally corrects the initial three-dimensional grid model based on real-time visual data to generate a high-precision dynamic three-dimensional grid model; The path planning module 25 is used to generate a cleaning path using an adaptive grid traversal algorithm; The obstacle detection and path re-planning module 26 is used to detect obstacles in the path in real time through a visual feature matching algorithm and to re-plan the path to avoid obstacles; The cleaning execution control module 27 is used to control the UAV body 21 to complete the high-altitude cleaning operation through the cleaning execution device 212 according to the planned path.

[0128] Finally: the above-mentioned only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A method for precise path planning for high-altitude cleaning by unmanned aerial vehicles (UAVs) based on vision navigation, characterized in that, Includes the following steps: Construct an initial 3D mesh model of the building to be cleaned, and divide the cleaning area according to the initial 3D mesh model and set the priority; The drone uses multimodal vision sensors to collect visual data of the cleaning area, including surface texture, depth information, and heat maps of stain distribution; The initial 3D mesh model is locally corrected based on real-time visual data to generate a high-precision dynamic 3D mesh model. An adaptive grid traversal algorithm is used to generate a cleaning path based on the priority of the cleaning area and the coverage of the drone cleaning head. The system uses a visual feature matching algorithm to detect obstacles in the path in real time. When an obstacle is detected, the path is replanned to avoid it until all areas are cleaned.

2. The method according to claim 1, characterized in that, When constructing the initial 3D mesh model of the building to be cleaned, the following steps are included: The RGB camera on the drone during pre-flight acquires color images of the building surface, and the depth camera acquires distance information. Combined with simultaneous localization and mapping algorithms, a rough 3D point cloud model is generated, which is then converted into an initial 3D mesh model through mesh simplification and texture mapping.

3. The method according to claim 2, characterized in that, When dividing the cleaning area and setting priorities based on the initial 3D mesh model, the following are included: Based on the surface geometry features of the initial 3D mesh model, the building surface is divided into multiple independent cleaning areas; Using pre-flight infrared thermal imaging data or color image texture analysis, the average stain distribution density of each area is initially assessed, and a priority score is calculated using a weighted formula, taking into account the vertical height and area size of the area. The cleaning priority order of each cleaning area is determined by sorting the priority scores from highest to lowest.

4. The method according to claim 3, characterized in that, When collecting visual data of the cleaning area, the following should be included: Color images of the cleaning area are captured by an RGB camera at a preset frame rate. Texture features are extracted from the images to generate a surface texture feature map that includes color distribution and texture complexity. The distance data of each point in the cleaning area relative to the drone is obtained in real time by a depth camera. Combined with the real-time pose information of the drone, the three-dimensional spatial depth coordinates of each point in the cleaning area are calculated. The heat radiation intensity data of the cleaning area is collected by an infrared thermal imaging module, and the heat radiation value is mapped into a heat map with different color gradients. At the same time, the heat map is corrected by combining the analysis results of the proportion of dark stains in the color images collected by the RGB camera, so as to obtain a stain distribution heat map that accurately reflects the distribution of stains.

5. The method according to claim 4, characterized in that, When generating a high-precision dynamic 3D mesh model, the following steps are included: The surface texture feature map, 3D spatial depth coordinates and stain distribution heat map collected in real time are converted into the unified coordinate system of the initial 3D mesh model using the real-time pose information of the UAV, to ensure data spatial consistency. By comparing the deviations between the geometric parameters of the corresponding region of the initial 3D mesh model and the 3D spatial depth coordinate data, as well as the detailed differences between the initial texture mapping and the surface texture feature map, regions with deviations exceeding a preset threshold are selected as regions to be corrected. For the area to be corrected, the positions of the mesh vertices are adjusted using three-dimensional spatial depth coordinate data, and the local mesh topology is optimized to match the actual geometry of the building surface. The initial texture mapping of the area to be corrected is replaced with a surface texture feature map to preserve the real surface details; at the same time, the high stain density areas in the stain distribution heatmap are marked as key areas of the dynamic 3D mesh model to enhance the accuracy of its geometric and texture details.

6. The method according to claim 5, characterized in that, When generating the cleaning path, the following are included: The current cleaning area is assigned according to the cleaning priority order of each cleaning area; The optimal traversal direction is determined based on the surface geometry of the cleaning area; Adjust the spacing and length of the path segments according to the coverage area of ​​the drone cleaning head, so that the overlap rate of adjacent path segments is controlled within 10%-15%; Bézier curves are used to fit the path inflection points, with priority given to paths that are continuous in the same direction.

7. The method according to claim 1, characterized in that, When detecting obstacles in a path in real time using a visual feature matching algorithm, the following are included: Local feature points are extracted from real-time frames of the drone's RGB camera, and the three-dimensional spatial depth coordinates corresponding to the local feature points are obtained simultaneously from the depth camera. The thermal radiation intensity value corresponding to the local feature points is obtained from the infrared thermal imaging module. The extracted local feature points are matched with the pre-stored feature library of the corresponding region in the dynamic 3D mesh model, and the positional deviation, depth deviation and thermal radiation deviation of the feature points are calculated. Preset position deviation threshold, depth deviation threshold, and thermal radiation deviation threshold. When any deviation exceeds the corresponding threshold, the area where the feature point is located is marked as a suspected obstacle area. For areas suspected of being obstacles, the presence of obstacles is confirmed by tracking and verifying the visual data of multiple consecutive frames. Then, the three-dimensional spatial depth coordinates, the size of the circumscribed rectangle, and the outline shape of the obstacle are calculated using depth camera data.

8. The method according to claim 7, characterized in that, When path replanning is performed upon obstacle detection, it includes: By using the obtained 3D spatial depth coordinates, circumscribed rectangle size and outline shape of the obstacle, combined with the geometric boundary of the corresponding region in the dynamic 3D mesh model, the range of influence of the obstacle is determined. The optimal strategy is selected based on the relative positional relationship between the obstacle's influence range and the current cleaning path, where: If the area affected by the obstacle is less than 30% of the current path segment length: adopt a temporary jump strategy, adjust the end position of the current path segment, skip the area affected by the obstacle, and then resume the original traversal direction; If the obstacle's influence range is greater than or equal to 30% of the current path segment length: adopt an avoidance strategy, based on the A* algorithm or the fast extended random tree algorithm, to search for the shortest avoidance path from the current position to the target position on the other side of the obstacle in the dynamic 3D mesh model, and restore the original traversal direction after the avoidance is completed.

9. The method according to claim 8, characterized in that, When searching for the shortest bypass path, the following are included: Based on the three-dimensional spatial depth coordinates, circumscribed rectangle size, and outline shape of the obstacles in the dynamic three-dimensional mesh model, and combined with the building structure boundaries preserved within the model, it is determined that the path must not cross the obstacle body and the building structure boundaries, and the minimum safe distance between the path and the obstacle is not less than 20% of the coverage area of ​​the drone cleaning head; Construct a multi-dimensional cost function, including: total path length, safe distance between the path and obstacles, rate of curvature change at path turning points, and degree of overlap between the path coverage area and high-stain density areas in the stain distribution heatmap; Using the A* algorithm or the fast expanding random tree algorithm, an initial bypass path is generated within the search space corresponding to the dynamic three-dimensional mesh model, with the current position as the starting point and the target position on the other side of the obstacle as the ending point, based on the multi-dimensional cost function. The initial bypass path is smoothed, and the spacing between path segments is adjusted to ensure that the overlap rate of adjacent path segments is maintained at 10%-15%. Verify whether the optimized path meets the constraints and cleaning coverage requirements.

10. A visual navigation-based UAV high-altitude cleaning path precision planning system, characterized in that, During runtime, the method of claim 1 is executed, comprising: The drone itself is equipped with a flight control unit and a cleaning execution device; The multimodal visual perception unit includes an RGB camera, a depth camera, and an infrared thermal imaging module, which are used to collect surface texture, depth information, and heat maps of stain distribution in the cleaning area. The initial model building module is used to generate the initial 3D mesh model, divide the cleaning area, and set the priority. The dynamic model update module performs local corrections on the initial 3D mesh model based on real-time visual data to generate a high-precision dynamic 3D mesh model. The path planning module is used to generate cleaning paths using an adaptive grid traversal algorithm. The obstacle detection and path replanning module is used to detect obstacles in the path in real time through a visual feature matching algorithm and to replan the path to avoid obstacles. The cleaning execution control module is used to control the UAV body to complete the high-altitude cleaning operation through the cleaning execution device according to the planned path.

Images