Unmanned aerial vehicle spraying system for interchange road accumulated snow recognition and path dynamic planning

Abstract

The invention discloses an unmanned aerial vehicle spraying system for interchange road accumulated snow recognition and path dynamic planning, and belongs to the technical field of intelligent traffic and municipal environmental sanitation. The system comprises a road accumulated snow recognition module, a CAD drawing analysis module, a take-off and landing point management module, a material and electric quantity estimation module, a dynamic task splitting module, a path generation and sorting module, a supply judgment module and a visualization module. Accumulated snow is recognized at the edge end in real time through a lightweight FLSD model, and spraying path dynamic planning based on accumulated snow distribution, resource state and safety margin is achieved in combination with a dynamic consumption model and a comprehensive cost function. The system supports task elastic splitting, intelligent homeward voyage and supply decision making, and provides full-process visual output. Compared with a traditional method, the precision, safety and resource utilization efficiency of snow removal of the complex interchange road are remarkably improved, and the method is suitable for intelligent winter maintenance of high-risk areas such as viaducts and ramps.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent transportation and municipal environmental sanitation, and specifically relates to a UAV spraying system for overpass snow identification and path dynamic planning, which integrates computer vision, edge computing, path planning and autonomous decision-making technology and is suitable for snow identification and accurate spraying operation in complex overpass scenarios. BACKGROUND

[0002] At present, road winter snow removal operation mainly relies on manual cleaning, salt spreading vehicles or large-scale mechanical snow removal equipment. Such traditional methods have the advantages of high operation efficiency and wide coverage in the main trunk road or open road section of the highway, but in the complex structure of the overpass, the intersection area of multiple ramps and the curved viaduct, etc. Significant limitations are exposed: on the one hand, large equipment is difficult to enter narrow or high areas, and there are operation blind areas and safety risks; on the other hand, the existing operation mode generally adopts fixed route or timing scheduling strategy, and lacks real-time sensing ability for actual snow distribution, which cannot realize "on-demand spraying", resulting in problems such as waste of snow melting agent, snow removal not in time or over operation, etc.

[0003] In recent years, some researches have tried to introduce visual technology to identify road snow, such as image analysis method based on vehicle-mounted or fixed camera. However, these schemes are mostly offline processing or only used for post-evaluation, and cannot be deeply integrated with mobile operation platforms such as UAVs, and more cannot form a closed-loop control of "perception - decision - execution". In addition, the existing UAV spraying system usually adopts preset flight route, and the path planning is static and fixed, which cannot dynamically adjust the task order and flight strategy according to the real-time snow level, remaining power or spraying material, resulting in low resource utilization, high task interruption risk, and difficulty in adapting to the actual needs of complex overpass topology and changing environment.

[0004] In summary, the existing technology needs to solve the following core problems: 1. How to realize real-time and high-precision identification of road snow under complex lighting and background interference on the UAV platform with limited computing power; 2. How to deeply couple snow perception results with task planning to build a dynamic and reconfigurable spraying path generation mechanism; 3. How to consider power, material consumption, return safety and operation efficiency in the flight process to realize intelligent task scheduling and flexible execution under resource constraints.

[0005] Therefore, there is an urgent need for a UAV spraying system that integrates multi-modal perception, edge intelligence, dynamic path planning and autonomous decision-making to improve the intelligence, precision and safety level of winter snow removal operation in complex overpass scenarios. SUMMARY

[0006] To solve the above problems, the present application provides an unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning, aiming to build an intelligent operation system integrating "real-time snow perception-dynamic task planning-resource adaptive scheduling-closed-loop execution feedback". The system realizes high-precision and low-delay identification of complex overpass surface snow area at the edge of the unmanned aerial vehicle through the fusion of infrared and visible light multi-modal visual perception module; combined with CAD paper automatic analysis to generate structured flight segments, and based on multi-dimensional constraints such as snow distribution, remaining power, spraying material and return safety margin, dynamically generates and optimizes the spraying path, supports flexible splitting and intelligent supply decision of the task. Thus, the operation goal of "on-demand response, accurate snow removal, resource optimization" is achieved, and the efficiency, safety and intelligent level of snow removal operation in complex road network scenes such as viaducts, ramps and curves are significantly improved.

[0007] The technical scheme adopted by the present application is: The unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning comprises: a road snow identification module, a path generation and sorting module, and a material and power estimation module. The road snow identification module realizes low-delay and high-precision real-time snow detection under complex lighting conditions at the edge of the unmanned aerial vehicle based on a lightweight FLSD model, and outputs the real-time snow detection result to the path generation and sorting module. The path generation and sorting module dynamically generates a continuous spraying path for execution based on the current starting position and the flight segment set, using a comprehensive cost function that integrates entry distance, return distance and resource safety margin, and an improved greedy algorithm. The material and power estimation module accurately predicts the demand for spraying materials and battery power in real time based on flight segment geometric parameters and environmental factors through a dynamically adjusted consumption model, providing resource constraint basis for task scheduling and supply decision.

[0008] Further, the road snow identification module is based on an improved YOLOv10 to build a lightweight model FLSD, which introduces depth separable convolution + CSP structure in the backbone network, introduces spatial attention and channel attention multi-scale feature fusion attention mechanism in the feature extraction stage, and uses structured pruning, INT8 weight quantization and operator fusion in the edge inference stage for real-time edge inference; The path generation and sorting module uses an improved greedy algorithm and introduces a comprehensive cost function: ; In the formula, is the comprehensive cost score of the candidate flight segment ; is a unvisited spraying flight segment to be evaluated. is the minimum entry distance from the current UAV position to the leg ; is the return distance from the position after completing the leg to the selected landing point; is the resource safety margin coefficient after executing the leg ; , , are the entry distance weight, return distance weight, and safety margin weight, respectively; The dynamic adjustment consumption model established by the material and power estimation module is: , ; In the formula, is the consumption of spraying material required for executing the current leg; is the unit distance material consumption coefficient, indicating the amount of material required for spraying per unit distance of flight by the UAV; is the horizontal projection length of the leg, i.e., the path length that the UAV needs to spray the work; is the battery power consumption required for executing the current leg; is the unit horizontal distance power consumption coefficient, reflecting the energy consumption efficiency of the UAV in the horizontal uniform flight state; is the unit climbing height power consumption coefficient, reflecting the additional energy consumption caused by the vertical climbing of the UAV; is the net climbing height of the leg, i.e., the difference between the terminal altitude and the starting altitude; coefficient , , According to the wind speed, air temperature, and load, these environmental factors are adjusted in real time.

[0009] Further, the UAV spraying system for identifying snow accumulation on overpass roads and dynamically planning paths further includes a CAD drawing analysis module that interacts with the path generation and sorting module; The CAD drawing analysis module extracts and standardizes the two-dimensional coordinate sequence of the road spraying leg by automatically reading the line graphic elements in the DXF file, generates a structured, non-redundant, and geometrically accurate leg dataset, and transmits it to the path generation and sorting module.

[0010] Further, the CAD drawing analysis module automatically analyzes the LINE / LWPOLYLINE graphic elements in the DXF format file, extracts the leg key coordinate point sequence, completes direction standardization, overlap merging, and boundary correction; defines the coordinate conversion function provides structured input for plane path planning; In the formula, A three-dimensional point in the original CAD drawing. or ,Include coordinate; It is a coordinate transformation function used to project or simplify a point in three-dimensional space into two-dimensional plane coordinates; Point The horizontal coordinate on the horizontal plane; Point The vertical coordinate on the horizontal plane; These are the transformed two-dimensional coordinates, representing the position of the point within the working plane; This represents the two-dimensional Euclidean space consisting of all pairs of real numbers, i.e., the Cartesian coordinate system, which is used as input for subsequent path planning algorithms.

[0011] Furthermore, the UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes a take-off and landing point management module that interacts with the material and power estimation module and the path generation and sorting module. The take-off and landing point management module introduces an accessibility scoring function that integrates distance, time, and resource safety margins. Under the premise of meeting the power and material safety thresholds, it selects the optimal return target from multiple candidate take-off and landing points.

[0012] Furthermore, the accessibility scoring function is: ; In the formula, Candidate take-off and landing points The overall accessibility score is as follows: the higher the value, the better. Candidate take-off and landing points; This indicates the current location of the drone. To start from the current location of the drone to candidate take-off and landing points The range of flight; The maximum return distance allowed for the mission; To fly from the current location Predicted flight time; The maximum allowable return flight time for the mission; Safety margin coefficient reflects the degree of resource redundancy; , , The dynamically adjustable weight parameters are for the distance, time, and safety margin terms, respectively, to satisfy... This is used to adapt to different task priorities; Safety margin coefficient Defined as: ; In the formula, The current remaining amount of the unmanned aerial vehicle; The take-off and landing point to be flown to The required return flight amount of electricity; The current remaining amount of spraying material of the unmanned aerial vehicle; The required remaining amount of material that may be consumed or reserved during the return flight; Only the take-off and landing point that meets And The take-off and landing point is scored preferably.

[0013] Further, the unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning further comprises a replenishment determination module interacting with the material and electricity estimation module and the take-off and landing point management module; The replenishment determination module constructs a dynamic threshold model based on the remaining resources, the return flight distance and the environmental factors, determines the replenishment demand in real time, records the triggering conditions and the replenishment types, and supports task tracing and system optimization.

[0014] Further, the unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning further comprises a dynamic task splitting module interacting with the path generation and sorting module; The dynamic task splitting module automatically interrupts and divides the remaining tasks during the flight segment execution based on real-time resource evaluation and safety margin determination, ensures that the operation does not fail in the middle due to depletion of electricity or material, and supports traceability and reconstruction of task status.

[0015] Further, the unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning further comprises an output and visualization module interacting with each module; The output and visualization module displays the whole process of the task through a graphical interface, generates a structured report containing flight, spraying and resource consumption data, and supports multi-format export for archiving and analysis.

[0016] The beneficial effects of the present application are: The unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning of the present application has significant technical advantages and practical value in complex overpass snow removal operations through multi-module cooperation and intelligent closed-loop design, and the specific beneficial effects are as follows: 1. Real-time snow perception with high precision and low delay: The lightweight FLSD model is adopted, which integrates deep separable convolution, CSP structure and multi-scale attention mechanism, and realizes 20ms / frame inference speed and high detection accuracy on the edge of the unmanned aerial vehicle, effectively dealing with complex road scenes such as strong reflection, shadow, background interference, etc., providing reliable perception basis for "on-demand spraying".

[0017] 2. Break through the limitations of traditional static path planning and achieve dynamic intelligent scheduling: Based on snow distribution, resource status and safety margin, the improved greedy algorithm is used to dynamically generate spraying paths, comprehensively optimize the entering distance, return feasibility and resource redundancy, avoid the risk of "short distance but no return", and significantly improve the path rationality and task continuity.

[0018] 3. Improve resource utilization efficiency and reduce operation and maintenance cost: Through the dynamic consumption model, the material and power demand is accurately predicted, and combined with the dynamic task splitting and supply judgment mechanism, invalid flight and excessive spraying are avoided, and material and energy waste is significantly reduced.

[0019] 4. Ensure operation safety and task reliability: Introduce the landing point accessibility scoring function and safety margin threshold screening mechanism to ensure that the UAV always operates within the safety boundary; automatically split and record the state when the task is interrupted, support breakpoint resume, and eliminate the risk of failure in the middle.

[0020] 5. Realize full-process automation and visual management: From automatic analysis of the flight segment from CAD drawings to snow identification, path planning, and execution feedback, no manual intervention is required; the output module provides a graphical task view and structured report, supports CSV / PDF export, and is convenient for scheduling supervision, performance evaluation and strategy optimization.

[0021] 6. Especially suitable for complex interchange scenarios, filling the technical gap: For high-risk areas that are difficult to cover by large equipment such as viaducts, multiple ramps and sharp turns, flexible, safe and efficient snow removal solutions are provided, significantly improving the safety and security of urban expressway network in winter.

[0022] 7. Good scalability and potential for engineering landing: The system architecture supports the expansion to multi-machine collaborative operation, real-time obstacle avoidance, weather linkage warning and other advanced functions, and the model and algorithm have been verified on embedded platforms such as Jetson Xavier NX, with actual deployment conditions.

[0023] In summary, the system not only solves the core pain points of existing snow removal technology in perception, planning and resource management, but also promotes the intelligent, fine and green upgrade of municipal winter maintenance, with outstanding technical progress and broad application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0025] Figure 1A schematic diagram of the composition of a drone spraying system for snow accumulation recognition and dynamic path planning on grade-separated roads; Figure 2 This is a schematic diagram of the overall architecture of the road snow accumulation recognition module; Figure 3 Image showing the effect of identifying snow accumulation on roads; Figure 4 A schematic diagram of the overall highway interchange and air route CAD drawing; Figure 5 This is a visual diagram of the flight path for mission segment 10. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] This embodiment provides a drone spraying system for snow accumulation recognition and dynamic path planning on grade-separated roads, aiming to solve the problems of perception lag, rigid path, resource waste and high operational risks faced by traditional road snow removal methods on complex grade-separated roads such as viaducts, multiple ramps and sharp bends.

[0028] like Figure 1 As shown, this UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads includes: a road snow accumulation recognition module, a path generation and sorting module, and a material and power estimation module. The specific details of each module are as follows: Road snow accumulation recognition module: The road snow recognition module is deployed on the UAV's onboard edge computing unit, such as NVIDIA Jetson XavierNX or a domestic embedded SoC. It is used to detect snow areas in real-time video streams collected by onboard visible light and infrared cameras before or during spraying tasks, and output structured perception results for the path generation and sorting module to dynamically adjust the operation strategy.

[0029] like Figure 2 As shown, the road snow recognition module, based on the improved YOLOv10 target detection network, constructs a fast and lightweight snow surface detection model, FLSD, the specific implementation of which is as follows: Step 1. Network structure lightweight: In the backbone network, introduce depthwise separable convolution, which decomposes the standard convolution into channel-wise convolution and 1x1 point-wise convolution, significantly reducing the amount of calculation and parameter; At the same time, adopt the cross-stage partial connection structure CSP, divide the input feature channel into the reserved path and the transformation path, reduce the redundant information transmission through cross-layer fusion. After optimization, the model size is reduced to about 60% of the original YOLOv10 backbone, the inference speed is increased by 25%, and the mAP precision is basically unchanged, effectively adapting to the computing power constraint of unmanned aerial vehicle embedded platform.

[0030] Step 2. Feature extraction enhancement: In view of the identification difficulties of road snow in strong light reflection, shadow shielding and complex background, such as ruts and vegetation, a multi-scale attention mechanism is introduced in the feature fusion stage, including: spatial attention module: weight each spatial position of the feature map, enhance the boundary discrimination ability of high reflection area and shadow area; Channel attention module: adaptively adjust the channel weight, highlight the snow surface texture and edge features, and suppress irrelevant interference; Multi-scale feature fusion structure: aggregate high, medium and low resolution features, considering large-area snow cover and small residual patch detection.

[0031] As shown in Figure 3 , the system conducts multiple comparative experiments in real road snow scenes, focusing on verifying the detection robustness of the FLSD model under complex lighting conditions such as backlight, snow mirror reflection, and high dynamic range background. The green rectangular frame in the figure is the snow area boundary successfully identified by the model. Although the image has strong sunlight reflection and shadow interference, the model can still accurately frame the snow coverage range, and is not affected by ruts or roadside vegetation.

[0032] The experiment is based on the self-built "overpass road snow dataset", covering morning backlight, afternoon strong reflection, dusk low illumination and sunny day after snow, etc. Compared with the original YOLOv10 model, the FLSD model with multi-scale attention mechanism has an average precision of 82.9% under the above complex conditions, an increase of about 4.7 percentage points compared with the original YOLOv10 model. At the same time, the false detection rate is reduced by 12.3%, and the missed detection rate is reduced by 8.6%, indicating that the mechanism effectively enhances the model's ability to distinguish high-brightness reflection areas and texture-fuzzy areas.

[0033] The results fully prove that the proposed multi-scale attention fusion structure can significantly improve the stability of snow perception in real complex environments for unmanned aerial vehicles, providing reliable input basis for subsequent dynamic path planning.

[0034] Step 3. Edge inference optimization: To meet the real-time requirements, post-training model deployment-level optimization is performed: structured pruning is used to remove redundant convolution kernels based on channel importance scores, reducing computational complexity to 65% of the original, while preserving key feature pathways; INT8 weight quantization is implemented to convert the FP32 model to INT8 format, which improves inference speed by 1.8-2.2 times on ARM NPU or GPU, with an accuracy loss of less than 1%; operator fusion and memory access optimization are combined with hardware characteristics to achieve pipeline parallel processing of video decoding, preprocessing, and inference, with an end-to-end delay of 20ms / frame, 50FPS, meeting real-time job requirements.

[0035] During task execution, this module outputs structured snow information every frame, including snow area bounding box coordinates, coverage ratio estimated based on pixel proportion, confidence score, etc. The above data is transmitted in real time to the path generation and sorting module for dynamic adjustment of spraying path order, flight speed or local operation mode, thereby significantly improving snow removal efficiency and resource utilization while ensuring job safety.

[0036] Path generation and sorting module: The path generation and sorting module is used to dynamically generate efficient, safe and continuous executable spraying paths within each spraying task based on the current starting position of the UAV and the structured segment set provided by the system. This module is deployed on the UAV's on-board task management unit or the ground control station, and can combine real-time sensing information and resource status to achieve flexible task scheduling.

[0037] The "structured segment set" is provided by the system in advance, and its source can be a manually delineated work area, road network data imported from GIS maps, or road sprayable segments defined by other means. Each segment is represented as a sequence of two-dimensional coordinate points, constituting the smallest work unit executable by the UAV.

[0038] The specific implementation process is as follows: Step 1. Input data preparation: the module receives the following key inputs: a set of segments to be sprayed, each defined by a starting point, an ending point and intermediate path points; real-time snow detection results from the road snow identification module for filtering of segments to be actually executed, such as spraying only on segments with snow coverage ratio above a threshold; current real-time position of the UAV ; preferred take-off and landing point ; remaining battery level and remaining material .

[0039] Step 2. Path strategy selection: The system supports multiple path planning strategies, including: "Start-point priority" mode: prefer to start work from the segment closest to the current drone position, suitable for fast response scenarios; "Sequence maintenance" mode: execute according to the preset logical sequence of segments, such as following the ramp, suitable for curved or sloped areas with high requirements for operation continuity; The strategy can be specified by the operator or automatically switched by the system according to the task urgency.

[0040] Step 3. Improved greedy path sorting algorithm execution: The module uses an improved greedy algorithm, combined with resource safety constraints, to iteratively generate segment access sequences according to the following steps: First, initialize the set of unvisited and sprayed segments ; Then, for each candidate segment , calculate its minimum entry distance , that is, the shortest flight distance from the current position to any entry point of the segment; Then, calculate the return distance to the preferred take-off point after executing the segment; Then, based on the material and power estimation model, predict the remaining resources after execution , and calculate the safety margin coefficient : ; In the formula, , is the remaining power and material after execution; , is the resource required for return; Then, build a comprehensive cost function: ; In the formula, is the comprehensive cost score of the candidate segment ; is the unvisited spraying segment to be evaluated at present; is the minimum entry distance from the current drone position to the segment ; is the return distance from the position after completing the segment to the selected take-off point; is the resource safety margin coefficient after executing the segment ; , , are the entry distance weight, return distance weight, and safety margin weight, , , > 0 is a dynamically adjustable weight coefficient, for example, when the power is less than 30%, automatically increase and to prioritize the safety of the return; Then, select the segment with the smallest value as the next task segment, add it to the path sequence, and remove it from ; Finally, update the current starting point to the end of the segment, repeat the above process until all the spraying segments are processed or the supply condition is triggered.

[0041] Step 4. Local optimization and continuity guarantee: Based on the greedy main body, the module can perform 2-opt exchange and other lightweight optimization on the local subsequence of the generated path, such as continuous 3 segments; thereby reducing unnecessary U-turns, improving path smoothness and operation efficiency.

[0042] Through the above mechanism, the path generation and sorting module realizes dynamic path planning based on general structured segment input, perception-driven, resource-constrained, and safety-prioritized, effectively avoiding the problems of "close distance but unable to return" or "resource depletion midway" in traditional methods, significantly improving the reliability and intelligence level of complex interchanges road snow removal operations.

[0043] Material and power estimation module: The material and power estimation module is deployed in the unmanned aerial vehicle on-board task management system, which is used to dynamically estimate the spraying material and battery power required for the current to-be-executed segment or task subset before and during task execution, providing key resource constraints for dynamic task splitting, path planning, and supply judgment.

[0044] The inputs of the material and power estimation module include: the geometric parameters of the segment: horizontal projection length and net climb height ; the current state parameters of the unmanned aerial vehicle: load, flight speed, spraying rate; real-time environmental data obtained through on-board sensors or weather API, such as wind speed, air temperature, and air density.

[0045] The material and power estimation module has a built-in dynamic resource consumption model, which is as follows: Material consumption model: spraying material demand is calculated by the following formula: ; Where, is the spraying material consumption required to execute the current segment; is the horizontal projection length of the segment, i.e. the path length that the unmanned aerial vehicle needs to spray the operation; The material consumption coefficient per unit distance represents the amount of material required to be sprayed by the drone per unit distance it flies. For example, if the nozzle flow rate is 0.5 L / min and the flight speed is 5 m / s, then... =0.0017L / m.

[0046] In actual operation, It can dynamically adjust according to environmental and task requirements: when a high snow cover ratio is detected or the temperature is below -5℃, it automatically increases. To increase the concentration of de-icing agent; when encountering crosswinds or headwinds, to compensate for drift loss, the concentration should be appropriately increased. The system supports presets by region. Values ​​are used to achieve differentiated and precise spraying.

[0047] Power consumption model: Battery power demand Calculate using the following formula: ; in, Battery power consumption required to complete the current flight segment; The power consumption coefficient per unit horizontal distance reflects the energy efficiency of the drone in horizontal uniform speed flight. The power consumption coefficient per unit altitude reflects the additional energy consumption caused by the drone's vertical climb; This is the net climb altitude of the flight segment, which is the difference between the altitude of the endpoint and the altitude of the starting point.

[0048] coefficient and It is not a fixed constant, but is adjusted in real time based on the following factors: low temperatures increase battery internal resistance and decrease efficiency. , Automatic adjustment, such as increasing by 15% at 0°C compared to 25°C. Headwinds increase drag. The weight increases linearly in segments according to wind speed. As materials are consumed, the overall weight of the drone decreases. , Gradually reduce power. Activate high-power mode during the climb phase. Dynamic enhancement; additional energy consumption is introduced when hovering or turning at low speed.

[0049] Assuming a certain flight segment length =300m, elevation gain =15m, current temperature -3℃, wind speed (headwind) 4m / s, remaining load 12kg. The system queries the correction factor for the current environment: =0.0021L / m, =0.85Wh / m, =12.3Wh / m. Therefore: =0.0021×300=0.63L, =0.85x300+12.3x15=255+184.5=439.5Wh, which is output to the dynamic task splitting module and the supply determination module in real time to determine whether the safety margin condition is met.

[0050] Through the above dynamic modeling and real-time correction mechanism, the material and power estimation module significantly improves the accuracy and environmental adaptability of resource estimation, effectively supporting intelligent decision-making and safe operation of the system in complex interchange road scenarios.

[0051] Based on the above road snow identification module, path generation and sorting module, material and power estimation module, the "perception → decision → resource constraint" integrated unmanned aerial vehicle intelligent spraying system is constructed, realizing the on-demand response, accurate snow removal and resource optimal winter road maintenance goal.

[0052] Further, as a preferred technical solution of the embodiment, as shown in Figure 1 The unmanned aerial vehicle spraying system for interchange road snow identification and path dynamic planning also includes a CAD drawing analysis module, which interacts with the path generation and sorting module to automatically convert the road design CAD drawings provided by the municipal or transportation department into structured and executable spraying segment data, providing high-precision geometric input for subsequent path planning.

[0053] As shown in Figure 4 The original CAD drawing processed by the CAD drawing analysis module is the overall road network structure of the high-speed interchange bridge, including multiple ramps, main roads and intersection nodes, and the graph elements are represented by colored lines to represent different functional road segments, such as green for main roads, purple for ramps, and yellow for connecting channels. The system automatically extracts the boundaries of the spraying operation segments from such drawings and performs standardized processing.

[0054] The specific implementation process is as follows: Step 1. Input and graph element recognition: The module receives a CAD drawing file in standard DXF format, which usually contains vector graph elements of road structures such as interchanges, ramps, main roads, etc. The system traverses all entities in the DXF file, focusing on identifying line graph elements that define sprayable areas, including straight line segments LINE and lightweight multi-segment lines LWPOLYLINE, and filtering out non-operation-related graph elements such as text, labels, and fills.

[0055] Step 2. Segment coordinate extraction and preprocessing: For each valid graph element, extract its vertex coordinate sequence. For LWPOLYLINE, get all control points; for LINE, consider it as a segment composed of two endpoints. All original coordinates are stored in the form of three-dimensional vectors, denoted as ; where is a three-dimensional space point in the original CAD drawing. for elevation information.

[0056] Step 3. Coordinate transformation and two-dimensional projection: Since the UAV spraying path planning is mainly carried out in the horizontal plane, the system defines a coordinate transformation function: ; wherein, is a coordinate transformation function for projecting or simplifying a three-dimensional space point into a two-dimensional plane coordinate; denotes the horizontal coordinate of point on the horizontal plane; denotes the vertical coordinate of point on the horizontal plane; is the transformed two-dimensional coordinate, representing the position of the point in the working plane; denotes a two-dimensional Euclidean space consisting of all real number pairs, i.e., a plane rectangular coordinate system, for the input of subsequent path planning algorithms.

[0057] Step 4. Geometric post-processing and data optimization: To ensure the accuracy and executability of the flight segment data, the module performs the following geometric optimization operations: direction standardization: unify the direction of all flight segments, such as sorting by traffic flow direction or clockwise, to avoid direction conflicts during path planning; overlapping flight segment merging: detect and merge adjacent line segments that are spatially overlapping or highly close, eliminating redundancy and reducing task fragmentation; boundary flight segment automatic correction: for breakpoints, misplacement or small gaps caused by mapping errors, use interpolation or nearest neighbor connection strategy for automatic repair to ensure continuous closure of flight segments.

[0058] Step 5. Structured flight segment dataset output: finally, the module generates a structured flight segment dataset, with each flight segment represented as an ordered list of two-dimensional coordinate points: ; This dataset is transmitted to the path generation and sorting module through an internal interface as the basic task space for dynamic path planning and snow-responsive spraying scheduling.

[0059] Through the above processing, the CAD drawing analysis module realizes the automatic bridging from engineering design drawings to intelligent UAV task instructions, significantly improving the deployment efficiency and operation accuracy of the system in complex interchange road scenarios, while avoiding subjective errors and time costs of manual route marking.

[0060] Further, as a preferred technical solution of the present embodiment, as Figure 1As shown, the overpass snow identification and path dynamic planning unmanned aerial vehicle spraying system also includes a take-off and landing point management module. The take-off and landing point management module closely interacts with the material and power estimation module and the path generation and sorting module, and is used to intelligently select the optimal return target in a multi-take-off and landing point deployment scenario, and ensure that the unmanned aerial vehicle always has safe and efficient return capability in the complex operation process.

[0061] When the system presets multiple candidate take-off and landing points, such as emergency landing pads, service areas or designated safety areas around the overpass, the take-off and landing point management module evaluates the accessibility of each candidate point in real time during the task execution, and makes dynamic decisions based on the resource state.

[0062] The specific implementation process is as follows: Step 1: Feasibility preliminary screening: for each candidate take-off and landing point , the take-off and landing point management module first calls the material and power estimation module to obtain the resources required for return, including: : the return power required to fly to . : the remaining amount of spraying material that needs to be reserved or consumed during the return; in some scenarios, no material is consumed during the return, and 0 can be set, but a minimum amount is usually reserved for safety.

[0063] At the same time, the current state of the unmanned aerial vehicle is obtained: : the remaining power; : the remaining material amount. A safety margin threshold , is introduced, and only when the following two conditions are met, the take-off and landing point is considered as a feasible candidate: and ; The take-off and landing points that do not pass the screening are directly excluded to avoid critical return risks.

[0064] Step 2: Accessibility comprehensive score: for all feasible candidate take-off and landing points, the accessibility score function is calculated: ; In the formula, is the accessibility comprehensive score of the candidate take-off and landing point , and the larger the value, the better; is the candidate take-off and landing point; is the current position of the unmanned aerial vehicle; is the distance of the journey from the current position of the unmanned aerial vehicle to the candidate take-off and landing point ; is the maximum return distance allowed by the task; is the return power required to fly to predicted time of flight; maximum return flight time allowed for the task; safety margin coefficient, reflecting the degree of resource redundancy; safety margin coefficient is defined as: ; wherein, is the current remaining power of the UAV; is the return power required for the UAV to fly to the take-off and landing point ; is the current remaining amount of spraying material of the UAV; is the required material margin that may be consumed or reserved during the return process; , , are respectively the dynamically adjustable weight parameters of the distance item, the time item and the safety margin item, satisfying , for adapting to different task priorities, for example: when returning urgently, increasing to preferentially select the take-off and landing point with the shortest time consumption; when performing regular operations, increasing to preferentially ensure resource safety.

[0065] Step 3: Optimal take-off and landing point selection and output: The module sorts all feasible candidate points according to from high to low, and selects the point with the highest score as the current optimal return target, and transmits the coordinates of the point to the path generation and sorting module in real time, for calculating the return distance and updating the comprehensive cost function.

[0066] Through the above mechanism, the take-off and landing point management module realizes intelligent optimization of multiple take-off and landing points under the premise of ensuring safety margin, significantly improves the robustness and task reliability of the UAV in complex interchange road long-time operation, and effectively avoids the risk of crash or task interruption caused by blind return.

[0067] Further, as a preferred technical solution of the embodiment, as shown in Figure 1 , the UAV spraying system facing interchange road snow identification and path dynamic planning further comprises a supply determination module. The supply determination module interacts with the material and power estimation module and the take-off and landing point management module in real time, for dynamically evaluating whether it is necessary to interrupt the current operation and return for supply during the task execution, so as to ensure the continuous, safe and efficient operation of the system.

[0068] The specific implementation process is as follows: ​Step 1. Dynamic Threshold Model Construction: The replenishment determination module constructs an adaptive threshold determination model based on multi-source inputs. The core inputs include: the drone's current remaining battery power. and remaining amount of sprayed material From the materials and power estimation module; flying to the current preferred takeoff and landing point. Required return battery power Material allowance Real-time environmental factors, such as wind speed, temperature, and air density, are obtained through airborne sensors or meteorological interfaces; mission phase information, such as whether it is in a high-priority ramp operation.

[0069] The system dynamically adjusts the replenishment trigger threshold based on the above factors, rather than using a fixed percentage. For example, when the wind speed increases or the temperature decreases, the energy consumption for returning to base increases, and the system automatically raises the power warning threshold. In remote operating areas far from the take-off and landing points, even if there are still plenty of resources, replenishment may be triggered earlier due to the long return distance. If it is detected that the next flight segment has severe snow accumulation and requires high-concentration spraying, the material threshold will be tightened accordingly.

[0070] Step 2. Supply condition judgment logic: The module continuously monitors whether any of the following conditions are met: ; ; in , To dynamically adjust the safety margin factor according to the environment and task, it is usually , Once any of the conditions are met, the return-to-base resupply command will be triggered immediately.

[0071] Step 3. Supply Event Recording and Type Labeling: When supply is triggered, the module automatically records a complete event log, including: trigger time; current task location; and remaining resource status. , The return destination point is from the takeoff and landing point management module. Supply type tags include, but are not limited to: "Low Power": triggered due to power falling below the dynamic threshold; "Materials Depleted": triggered due to insufficient snow melting agent reserves; "Task Segment Completed + Resource Critical": the current subtask has ended and resources are nearing the safety boundary; "Sudden Environmental Change": such as a sudden surge in energy consumption caused by strong winds. Logs are stored in a structured format, such as JSON or database records, and synchronized to the output and visualization module for generating task reports.

[0072] Step 4. Coordination with Task Scheduling: The resupply determination result is notified to the dynamic task splitting module in real time. Based on this, the module divides the incomplete flight segment into new tasks, ensuring that the operation can be interrupted and resumed. At the same time, after receiving the resupply instruction, the path generation module immediately plans the return route, prioritizing a safe landing.

[0073] Through the above mechanism, the replenishment determination module realizes the transition from "passive low power return" to "active predictive replenishment", not only avoiding the risk of failure during the task, but also providing data support for long-term operation of the system, supporting continuous optimization of the consumption model, path strategy and take-off point layout.

[0074] Further, as a preferred technical solution of the embodiment, as shown in Figure 1 The unmanned aerial vehicle spraying system facing the overpass snow identification and path dynamic planning further includes a dynamic task splitting module. The module closely interacts with the path generation and sorting module and relies on the resource prediction data provided by the material and power estimation module to realize intelligent and safe job interruption and task reconstruction during task execution, ensuring that the unmanned aerial vehicle will not be forced to stop in the middle of the flight segment due to power or spraying material depletion.

[0075] The specific implementation process is as follows: Step 1. Triggering opportunity and evaluation mechanism: the system starts the task continuity evaluation process immediately after each flight segment is executed, i.e. when the unmanned aerial vehicle completes a spraying job and arrives at the end of the flight segment. The module calls the material and power estimation module to obtain the predicted resource consumption value required for executing the next to-be-visited flight segment , , , and makes a judgment in combination with the current remaining resources , .

[0076] Step 2. Safety margin judgment logic: a dynamic safety margin threshold , is introduced. If any of the following conditions is met, it is determined that the next flight segment cannot be safely completed, and the task splitting is triggered. ; ; This mechanism ensures that even if the environment mutates, such as increased adverse wind or there is a small error in the model, the system still has enough redundancy to complete the current flight segment and safely return.

[0077] Step 3. Task interruption and splitting operation: interrupt the current task flow, send the "terminate spraying, prepare for return" instruction to the flight control system; encapsulate the remaining unexecuted flight segment set, including and subsequent flight segments, as a new subtask and assign it a unique task ID; update the task state database, mark the original task as "partially completed" and the new subtask as "to be dispatched"; notify the path generation and sorting module to re-plan the spraying path of the subtask based on the new starting point before the next take-off.

[0078] Step 4. Task state logging and traceability: The module synchronously generates a structured task segmentation log, recording the following key information: original task ID and new subtask ID; time, location, and completed leg list when splitting occurs; splitting reason, such as "low battery" or "material critical"; remaining resource status at the time of splitting and prediction basis. The log supports complete traceability, visual playback, and fault review of task status, and is used for task reconstruction when the system resumes work, i.e., seamless continuation of flight from the breakpoint, avoiding repeated spraying or missing areas.

[0079] Step 5. Coordination with replenishment process: Dynamic task splitting module and replenishment decision module linkage: if the splitting is caused by resource shortage, the system automatically enters the replenishment process; after the replenishment is completed, the scheduler loads the subtask first, realizing the "interruption-replenishment-continuation" closed loop.

[0080] Through the above mechanism, the dynamic task splitting module significantly improves the robustness and reliability of the system in complex and long-time work scenarios, truly realizing the intelligent work capability of "task interruptible, state savable, work continuable", effectively solving the technical bottleneck of task failure caused by resource depletion in traditional unmanned aerial vehicle spraying systems.

[0081] Further, as a preferred technical solution of the embodiment, as shown in Figure 1 , the unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning further comprises an output and visualization module. The output and visualization module serves as the human-machine interaction and data archiving hub of the system, and interacts in real time with the road snow identification module, CAD drawing analysis module, path generation and sorting module, dynamic task splitting module, replenishment decision module, material and power estimation module, and other functional units, aggregates task lifecycle data, and provides intuitive graphical display and structured report output.

[0082] The specific implementation process is as follows: Step 1. Multi-source data aggregation: The module continuously receives key data streams from various submodules during task execution, including: original CAD drawings and structured leg sets after analysis from the CAD drawing analysis module; real-time snow detection results, such as snow area heat map and coverage ratio, from the road snow identification module; dynamically generated spraying paths, leg access order, and return trajectories from the path generation and sorting module; task splitting records and replenishment event logs, including trigger time, location, and reason from the dynamic task splitting module and replenishment decision module; resource data such as flight distance, flight time, spraying area, material consumption , and power consumption from the material and power estimation module and flight log for each leg.

[0083] Step 2. Graphical interface display: Based on OpenCV, Matplotlib, or lightweight GUI frameworks, the module builds a unified visualization interface to dynamically present the following content: Background map overlay: Optionally load satellite maps or CAD base maps; Flight segment distribution: Identify completed, pending, and skipped snow-free segments with different colors; Drone trajectory: Real-time drawing of flight path, marking current position and heading; Key node markers: Mark take-off and landing points, supply points, and task splitting locations with icons; Resource status curve: Simultaneously display the trends of power and material changes over time / distance; Snow response heat map: Map the identification results to the road surface to intuitively reflect the "on-demand spraying" effect.

[0084] All graphical elements support zooming, panning, and layer switching to facilitate real-time monitoring and post-mortem analysis by dispatch personnel.

[0085] Step 3. Structured task report generation: After the task is completed, the module automatically generates a detailed task report, including but not limited to the following content: Task basic information: Task ID, execution date, and work area; Flight segment execution details: Each segment number, start / end coordinates, length, flight time, spraying area, material usage, and power consumption; Resource summary: Total flight distance, total spraying volume, total power consumption, and material utilization rate; Exception event records: Number of supplies, number of task splits, trigger causes, and handling results; Performance indicators: Total path length, number of return flights, and work continuity score.

[0086] Step 4. Multi-format export and archiving support: The report supports export to multiple standard formats: CSV format: Easy to import into Excel or databases for statistical analysis; PDF format: Generate formal documents with charts for archiving, reporting, or auditing; JSON / XML format: For upper-level intelligent municipal platforms to call and achieve data interconnection.

[0087] All exported files are automatically named by task ID and stored in local or cloud-specified directories, supporting retrieval by time, area, and UAV number.

[0088] Through the above design, the output and visualization module not only improves the operability and transparency of the system, but also provides a high-quality data basis for operation optimization, strategy iteration and performance evaluation, which is an important part of the "intelligent perception - autonomous decision - closed-loop feedback" technical system.

[0089] In addition, in order to verify the unmanned aerial vehicle spraying system for overpass snow identification and path dynamic planning, the embodiment also carries out the following actual verification: As Figure 4 shown, Figure 4 is a high-speed overpass whole and flight segment CAD drawing analysis effect diagram. The system successfully extracts and standardizes the flight segment coordinate sequence of multiple ramps, main roads and connecting channels from the original DXF drawing, completes the direction unification, overlap merging and boundary correction. The colored lines in the figure represent different functional areas, such as green for main road, purple for ramp, and gray dashed line for auxiliary connection path generated by the system, and white "+" mark for candidate take-off and landing point. After manual comparison, the geometric accuracy error of the flight segment is less than 0.5 meters, which meets the requirements of subsequent path planning and spraying operation.

[0090] As Figure 5 shown, Figure 5 is a flight path visualization diagram of task segment 10. The task contains multiple consecutive flight segments, and the system generates the optimal access sequence based on the current unmanned aerial vehicle position and snow distribution information using an improved greedy algorithm. The blue solid line in the figure is the actual flight trajectory, the dashed line is the return path, and the red circle points "start point 1" and "end point 2" are the task start and end positions respectively. The data shown in the figure shows that the total flight distance is 1183.76 meters, the total time consumption is 4.47 minutes, and the material consumption is 76.65 kilograms. The path planning effectively avoids repeated return, and dynamically inserts a return node during execution according to the resource state, ensuring the integrity of the task.

[0091] Through the above field deployment and data verification, the system realizes: CAD drawing → automatic conversion of structured flight segment; snow perception → closed-loop response of path dynamic adjustment; resource constraints → intelligent decision of task splitting and supply; task whole process → complete support for graphical visualization and data archiving.

[0092] The experimental results show that the system can significantly improve the snow removal operation efficiency, reduce the path length by about 18% compared with the traditional fixed flight line, and reduce the material waste by more than 15%, without any mid-flight power failure or crash event; Therefore, the system has good engineering practical value and popularization prospect.

[0093] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A drone spraying system for snow accumulation recognition and dynamic path planning at grade-separated roads, characterized in that: The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads includes: a road snow accumulation recognition module, a path generation and sorting module, and a material and power estimation module; The road snow recognition module is based on a lightweight FLSD model, which enables low-latency and high-precision real-time snow detection under complex lighting conditions at the edge of the drone, and outputs the real-time snow detection results to the path generation and sorting module. The path generation and sorting module uses a comprehensive cost function that integrates entry distance, return distance and resource safety margin, based on the current starting position and the set of flight segments, and dynamically generates continuously executed spraying paths through an improved greedy algorithm. The materials and power estimation module, based on flight segment geometry parameters and environmental factors, uses a dynamically adjusted consumption model to accurately predict the demand for spraying materials and battery power in real time, providing a resource constraint basis for mission scheduling and resupply decisions.

2. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 1, characterized in that: The road snow recognition module is based on the lightweight model FLSD built on the improved YOLOv10. It introduces a depthwise separable convolution + CSP structure in the backbone network, introduces an attention mechanism of spatial attention and channel attention multi-scale feature fusion in the feature extraction stage, and adopts structured pruning, INT8 weight quantization and operator fusion for real-time edge inference in the edge inference stage. The path generation and sorting module employs an improved greedy algorithm, introducing a comprehensive cost function: ； In the formula, Candidate flight segments Overall cost score; This is an unvisited spraying segment currently awaiting evaluation; From the current drone position to the flight segment Minimum entry distance; To complete the voyage The return distance from the current position to the selected take-off and landing point; For the execution of the flight segment The subsequent resource security margin coefficient; , , These are, respectively, the entry distance weight, the return distance weight, and the safety margin weight; The dynamic consumption adjustment model established by the material and power estimation module is as follows: , ； In the formula, The amount of spraying material required to complete the current flight segment; The material consumption coefficient per unit distance indicates the amount of material that the drone needs to spray per unit distance it flies. The horizontal projection length of the flight segment is the path length that the drone needs to spray. Battery power consumption required to complete the current flight segment; The power consumption coefficient per unit horizontal distance reflects the energy efficiency of the drone in horizontal uniform speed flight. The power consumption coefficient per unit altitude reflects the additional energy consumption caused by the drone's vertical climb; The net climb altitude of the flight segment is the difference between the altitude of the endpoint and the altitude of the starting point. coefficient , , Adjustments are made in real time based on environmental factors such as wind speed, temperature, and load.

3. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 1, characterized in that: The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes a CAD drawing parsing module that interacts with the path generation and sorting module. The CAD drawing parsing module automatically reads the line type primitives in the DXF file, extracts and standardizes the two-dimensional coordinate sequence of the road spraying segment, generates a structured, non-redundant and geometrically accurate segment dataset, and transmits it to the path generation and sorting module.

4. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 3, characterized in that: The CAD drawing parsing module automatically parses LINE / LWPOLYLINE primitives in DXF format files, extracts key coordinate point sequences for flight segments, and performs direction standardization, overlap merging, and boundary correction; it also defines coordinate transformation functions. Provides structured input for planar path planning; In the formula, A three-dimensional point in the original CAD drawing. or ,Include coordinate; It is a coordinate transformation function used to project or simplify a point in three-dimensional space into two-dimensional plane coordinates; Point The horizontal coordinate on the horizontal plane; Point The vertical coordinate on the horizontal plane; These are the transformed two-dimensional coordinates, representing the position of the point within the working plane; This represents the two-dimensional Euclidean space consisting of all pairs of real numbers, i.e., the Cartesian coordinate system, which is used as input for subsequent path planning algorithms.

5. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 1, characterized in that: The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes a take-off and landing point management module that interacts with the material and power estimation module and the path generation and sorting module. The take-off and landing point management module introduces an accessibility scoring function that integrates distance, time, and resource safety margins. Under the premise of meeting the power and material safety thresholds, it selects the optimal return target from multiple candidate take-off and landing points.

6. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 5, characterized in that: The accessibility scoring function is: ； In the formula, Candidate take-off and landing points The overall accessibility score is as follows: the higher the value, the better. Candidate take-off and landing points; This indicates the current location of the drone. To start from the current location of the drone to candidate take-off and landing points The range of flight; The maximum return distance allowed for the mission; To fly from the current location Predicted flight time; The maximum allowable return flight time for the mission; Safety margin coefficient reflects the degree of resource redundancy; , , The dynamically adjustable weight parameters are for the distance, time, and safety margin terms, respectively, to satisfy... This is used to adapt to different task priorities; Safety margin coefficient Defined as: ； In the formula, This indicates the drone's current remaining battery power. To fly to the take-off and landing point Required return battery power; This represents the current remaining amount of material to be sprayed by the drone. To reserve or retain the necessary materials that may be consumed or needed during the return journey; Only for those who meet the requirements and The take-off and landing points are scored and optimized.

7. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 5, characterized in that: The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes a resupply determination module that interacts with the material and power estimation module and the take-off and landing point management module. The resupply determination module constructs a dynamic threshold model based on remaining resources, return distance, and environmental factors to determine the resupply needs in real time, and records the triggering conditions and resupply types, supporting task traceability and system optimization.

8. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 1, characterized in that: The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes a dynamic task splitting module that interacts with the path generation and sorting module. The dynamic task splitting module is based on real-time resource assessment and safety margin judgment. It automatically interrupts and divides the remaining tasks during the execution of a flight segment, ensuring that the operation does not fail midway due to power or material depletion, and supports the traceability and reconstruction of task status.

9. The UAV spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads according to claim 1, characterized in that: The drone spraying system for snow accumulation recognition and path dynamic planning on grade-separated roads also includes an output and visualization module that interacts with each module. The output and visualization module displays the entire mission process through a graphical interface and generates structured reports containing data on flight, spraying, and resource consumption. It supports export in multiple formats for archiving and analysis.

Images