Intelligent recommendation method for road construction based on drone inspection and deep learning algorithm

Abstract

The present invention discloses an intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, including: collecting the center pile coordinates, road width, and curve element information of the highway, drawing a two-dimensional plane map and generating a grid map; using drones to shoot rolling construction road surface images, and then obtaining the gradation information of the rockfill material corresponding to the rolling area through a deep learning algorithm, and finally predicting the density of the rockfill body through a continuous grading fractal model; through training and evaluation, selecting the optimal model to conduct real-time evaluation of the road surface density; according to the evaluation results, planning the driving path of the roller, and repeatedly rolling the area with insufficient density until the construction requirements are met; during the construction process, updating the image data set in real time and optimizing the model to improve the evaluation accuracy. The present invention monitors the rolling conditions of the construction site in real time through drones and guides the roller to complete the rolling task based on the deep learning algorithm and the path planning algorithm, which effectively improves the construction efficiency and compaction quality.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring and decision support for road construction, and relates to an intelligent assessment technology for road construction quality based on drone inspection and deep learning algorithms. Specifically, the present invention relates to a method for assessing road construction quality by using a high-definition camera carried by a drone to collect road image data in real time and analyzing the image through a deep learning algorithm. Background Art

[0002] Road construction is an indispensable part of urban infrastructure construction and plays a vital role in the convenience of urban transportation and the sustainable development of the economy. The precise formulation and optimization of construction plans are the decisive factors in ensuring construction quality and improving construction efficiency. Traditionally, the design and evaluation of construction plans mainly rely on manual experience and manual calculations. This method lacks scientificity and objectivity and is difficult to adapt to the complexity and changing conditions of the construction environment. In order to enhance the level of intelligence in road construction, modern information technology, including deep learning models, artificial intelligence, big data, etc., has been introduced to assist in the generation and evaluation of construction plans, thereby providing more reasonable and efficient construction suggestions.

[0003] Deep learning model technology uses advanced algorithms to analyze and learn large amounts of data, providing an intelligent analysis and prediction tool for road construction. This technology can process and analyze high-resolution image data collected by drone inspections, automatically identify road surface defects, uneven distribution of construction materials, insufficient compaction, and other problems, thereby achieving real-time monitoring and early warning of construction quality. The deep learning model can continuously learn from new data, adapt to changes in road conditions during construction, and provide scientific construction suggestions for the construction team.

[0004] Path planning technology combines optimization algorithms and geographic information systems to provide a method for finding and generating optimal or suboptimal driving routes for road construction. It can generate different driving routes based on the actual needs and conditions of the construction, such as the shape, width, slope, etc. of the road, and enable construction workers and managers to easily select and execute construction routes through maps and navigation systems.

[0005] The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm combines the efficient data collection capability of drone technology and the intelligent analysis capability of deep learning model. It analyzes the image data collected by drone through deep learning model and provides intelligent decision support for construction. This method takes advantage of the high accuracy and adaptability of deep learning model, enabling construction personnel and managers to make more reasonable decisions based on accurate data analysis, while improving the automation and intelligence level of construction, and significantly improving the quality and efficiency of road construction. Summary of the invention

[0006] Purpose of the invention: In view of the problems existing in the selection and optimization of road construction plans, such as the lack of manual experience and calculations, the complexity and variability of construction environments and conditions, etc., an intelligent recommendation method for road construction based on drone inspection and deep learning algorithms is provided. Through the most advanced deep learning and generation of optimal construction paths, intelligent decision support is provided for road construction. According to the actual needs and conditions of road construction, this method can use drones to shoot images of rolling construction pavement, and then obtain the gradation information of the rockfill material in the corresponding rolling area through a deep learning algorithm, and finally predict the density of the rockfill body through a continuous grading fractal model. Through training and evaluation, the optimal model is selected to conduct real-time evaluation of the pavement density. According to the evaluation results, the driving path of the roller is planned, and the areas with insufficient density are repeatedly rolled until the construction requirements are met. During the construction process, the image data set is updated in real time and the model is optimized to improve the evaluation accuracy.

[0007] Technical solution: To achieve the above purpose, the present invention provides an intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, comprising the following steps:

[0008] S1: Collect information such as the center stake coordinates, road width, curve elements, etc. of the highway, draw a two-dimensional plane map of the corresponding section of the highway and generate the corresponding grid map, while ensuring the real-time update of the grid map during the construction process;

[0009] S2: Use the Road Surface Classification Dataset (RSCD) released by Tsinghua University to train deep learning models, use indicators such as precision (P), recall (R), F1-Score, and mean average precision (mAP) to compare and verify models, continuously optimize model parameters, and select the deep learning model with the best effect;

[0010] S3: Use a high-resolution camera on a drone to capture images of the construction road surface, and use the best deep learning model to analyze the images, collect the gradation information of the road surface rockfill, and predict the density of the corresponding rolling area based on the fractal model of continuous gradation. Update the color of the corresponding area on the two-dimensional plane according to the density of different areas to visualize the rolling conditions of each area;

[0011] S4: Based on the density assessment results, determine which areas have insufficient density and need further compaction. If there are areas with insufficient density, calculate the shortest driving path for each roller (paver) to the area that does not meet the density requirement. Plan the driving path of each roller (paver) according to the principle of shortest path and largest coverage of compaction area.

[0012] S5: If the density of the area meets the requirements, continue to evaluate and compact the next area. If the density does not meet the requirements, repeat step S4 until the density of all areas meets the construction standards.

[0013] S6: During the entire construction process, the pavement image dataset is updated in real time, and the deep learning model is retrained and optimized based on the new data to improve the accuracy and reliability of the compactness assessment;

[0014] S7: Determine whether the compaction task of the section has been completed. If not, repeat steps S4 to S6. If the compaction task has been completed and the density of all areas meets the requirements, the work is completed, and the evaluation results provide a basis for construction quality control.

[0015] Furthermore, the method for drawing the two-dimensional plane map and the grid map of the corresponding section of the highway in step S1 is:

[0016] A1: Based on the curve elements and other information collected by S1, the direction and position of the road centerline, as well as the horizontal and vertical curve elements of the road, such as straight lines, circular curves, transition curves, etc., can be determined. The expression formula of the transition curve is as follows:

[0017]

[0018] Among them, s is the arc length on the transition curve, A is the transition curve parameter, and L is the transition curve length. These parameters can be obtained from the curve element;

[0019] A2: Based on the elements of the road centerline, calculate the coordinates and mileage of each point, as well as the tangent angle, curvature radius, curve length and other parameters of the curve to form a stake-by-stake coordinate table. Generally speaking, the content of the stake-by-stake coordinate table includes the following parts: stake number, coordinates, elevation, tangent angle, curvature radius;

[0020] A3: Use professional mapping software such as CASS to draw the plan view and longitudinal section view of the road centerline, as well as the graphics of the road sideline, roadbed, pavement and other parts according to the pile-by-pile coordinate table, convert the numerical data of the road into graphic data, and finally output a two-dimensional plan view;

[0021] A4: The drawn two-dimensional plane map is divided into square cells with a side length of 5 cm. Each cell is called a grid. Each grid represents an area in the construction section and can be used to represent obstacles, drivable space and other information. When the grid map is constructed, each roller (paver) is regarded as an observation point to facilitate the subsequent calculation of a reasonable construction path.

[0022] Furthermore, the method for generating the grid map in step A4 is:

[0023] B1: The observation data of the construction site is obtained through the sensors installed on the roller (paving machine), that is, the geographic coordinates and direction of the roller (paving machine), and the position of each observation point in the corresponding coordinate system and the corresponding grid number are converted according to the geographic coordinates;

[0024] B2: For the observation point, determine whether it is about to touch the obstacle and which grids it has passed along the way based on its distance and angle. For the grid that is about to touch the obstacle, update its status to Occupied, for the grid that has passed, update its status to Free, and for the grid that has not been observed, keep its status as Unknown;

[0025] B3: In order to deal with the noise and uncertainty of the sensor, as well as the dynamic changes of the environment, a probabilistic method is usually used to update the state of the grid;

[0026] B4: By repeating steps B2 to B3, a grid map of the environment can be gradually constructed and updated according to real-time observation data. Finally, it can be determined whether it is an obstacle or free space based on the state of the grid.

[0027] Furthermore, the specific method for updating the grid map state in step B3 is:

[0028] Occupancy Grid Map, which uses the Bayesian rule to calculate the probability of each grid being occupied, is iteratively updated based on observed data and prior knowledge. The update formula of the occupancy grid map is as follows:

[0029]

[0030] Among them, P(m i ∣z 1:t ,x 1:t ) represents the grid state sequence z at all given time steps t 1:t and all observation sequences x at time step t 1:t Under the condition of i The probability of being occupied, z 1:t represents the grid state sequence from time step 1 to t, x 1:t represents the observed data sequence from time step 1 to t;

[0031] In order to simplify the calculation, the log-likelihood ratio is usually used to represent the state of the grid, that is:

[0032]

[0033] In this way, the state of the grid can be represented by a real number instead of a probability value; the larger the state of the grid, the more likely it is to be occupied; a grid state of 0 means that the probability of the grid being idle and occupied is equal or unknown;

[0034] The benefit of using the log-likelihood ratio is that the update formula can be simplified to:

[0035] l i = l i +l occu (z t ,x t )-l 0

[0036] In this way, each time the state of the grid is updated, it is only necessary to add a constant to the current log-likelihood ratio to obtain the latest log-likelihood ratio without considering the previous observation data and the position of the roller (paver); this can greatly improve the computational efficiency and save storage space.

[0037] Furthermore, in the intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, the specific process of step S2 is as follows:

[0038] C1: Preprocess the Road Surface Classification Dataset (RSCD), which is required to include the gradation information of rockfill materials and label it;

[0039] C2: Divide the preprocessed data set into training set, validation set, and test set to ensure that the model can be learned and validated on data at different stages;

[0040] C3: Select a suitable deep learning model, such as Faster-RCNN or YOLO series algorithms, and train the model on the training set, and use the validation set to adjust and optimize the model parameters;

[0041] C4: Use the test set to evaluate the performance of the trained model. Use evaluation indicators such as precision (P), recall (R), F1-Score, and mean average precision (mAP) to comprehensively compare and verify the performance of the model. After comparison, we found that Faster-RCNN has better performance.

[0042] C5: Based on the evaluation results, iteratively optimize the model, which may include adjusting the network structure, improving the training strategy, and introducing regularization methods to improve the accuracy and robustness of the model;

[0043] C6: During the iterative optimization process, the validation set is continuously used to evaluate the model performance to ensure that the model does not overfit to the training data while maintaining high accuracy;

[0044] C7: Faster-RCNN was finally selected as the model for collecting rockfill gradation information in the compaction area, and was validated on the test set to ensure its effectiveness and reliability in practical applications.

[0045] Furthermore, in the intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, the specific process of step S3 is as follows:

[0046] D1: Integrate high-precision positioning systems such as the Global Positioning System (GPS) and Inertial Measurement Unit (IMU) on the drone to ensure that the road images captured accurately correspond to the actual geographic location and record the flight path and shooting posture of the drone;

[0047] D2: The high-resolution camera on the drone collects high-definition images of the construction road surface according to the preset flight altitude and speed, ensuring that the image coverage is comprehensive and has sufficient resolution;

[0048] D3: The collected image data is transmitted to a server or cloud platform in real time through a wireless communication module, such as Wi-Fi or Bluetooth, for further processing and analysis;

[0049] D4: The server uses the Faster-RCNN model to extract the gradation information of the rockfill material from the transmitted image data, and then uses the continuous gradation fractal model to process the gradation information of the rockfill material and evaluate the density, thus realizing intelligent evaluation of the pavement compaction quality.

[0050] D5: After collecting the density corresponding to each area, assign corresponding colors to different areas according to the density.

[0051] Furthermore, in the method of using drones and deep learning models to analyze construction road surface images, predict and visualize density, the specific method of evaluating density in step D4 is:

[0052] E1: According to the rockfill grading information collected by D4 and the fractal theory, the particle size distribution and particle mass distribution of the rockfill can be described by the fractal dimension D. The particle number N is related to the maximum particle size d max The relationship between can be expressed as a fractal distribution formula:

[0053]

[0054] Among them, N 0 is a constant, D is the fractal dimension, and d is the diameter of the particle;

[0055] Similarly, the particle mass P and the maximum particle size d max The relationship between can also be expressed as a fractal distribution formula:

[0056]

[0057] Among them, C 0 is a normalization constant that ensures that when d = d max When , P(d)=1;

[0058] In actual engineering, grading has a significant impact on the density of rockfill materials; the fractal dimension D of grading can reflect the uniformity of particle size distribution, which in turn affects the density after compaction; when the fractal dimension D is low, it means that the particle distribution is relatively uneven, there may be large gaps, resulting in lower density; when the fractal dimension D is high, it means that the particle distribution is relatively uniform, the gaps between particles are small, and a higher density can be achieved;

[0059] E2: Studies have shown that there is an optimal range of fractal dimension D within which the rockfill material can achieve maximum density; this range is usually between 2.22 and 2.63; within this range, the particle size distribution can achieve optimal inter-particle filling, thereby obtaining the best compaction effect;

[0060] E3: The density ρ can be expressed by the following formula:

[0061]

[0062] Where m is the mass of the rockfill in dry state, V total is the total volume of the rockfill, V voids is the volume of pores inside the rockfill, P is the porosity, K particle is the particle distribution coefficient, K water is the moisture content coefficient;

[0063] By adjusting the gradation of rockfill materials, which includes the number of particles N and the particle mass P, the fractal dimension D can be indirectly optimized, thereby improving the density; the optimization can be achieved by adjusting the ratio of particles of different particle sizes to achieve the best inter-particle filling effect.

[0064] Furthermore, in step D5, after collecting the density corresponding to each area, the specific method of assigning corresponding colors to different areas according to the density is:

[0065] F1: First, we use image segmentation techniques to separate different regions in the image, which represent different rolling stages or pavement materials;

[0066] F2: Next, we identify each separated area and determine the road surface condition they represent based on the compaction level; this process includes identifying whether the road surface has been compacted, the degree of compaction, and whether there are areas that need further treatment;

[0067] F3: After identification, we will merge regions with similar characteristics. All regions identified as "compacted" can be merged together for clearer representation in the visualization.

[0068] F4: Finally, we merge the merged area with the original image and update the color of the corresponding area on the 2D plane. This step uses a specific algorithm to adjust the color in the image to reflect the compaction condition of each area; the color update can intuitively reflect the compaction quality of the road surface.

[0069] Furthermore, the specific method of using a specific algorithm to update the color in step F4 is:

[0070] G1: The number and level of colors, as well as the corresponding relationship of colors, can be determined according to the different compaction qualities. For example, the compaction quality can be divided into three levels: "uncompacted", "partially compacted" and "fully compacted", corresponding to the colors red, light green and dark green respectively;

[0071] G2: Determine the boundaries of the dyeing area according to the Bresenham algorithm, determine the dyeing points through the bidirectional scanning algorithm, and then dye each area.

[0072] Furthermore, in step G2, the boundary of the dyeing area is determined according to the Bresenham algorithm, and the specific method of determining the dyeing point by the bidirectional scanning algorithm is:

[0073] The formulas of the Bresenham algorithm for determining the boundaries of the dyed area and the bidirectional scanning algorithm for determining the dyed points are as follows:

[0074] Assume the screen coordinates of the three vertices of the triangle are (x1, y1), (x2, y2), (x3, y3), then:

[0075] Δx 12 =x 2 -x 1

[0076] Δy 12 =y 2 -y 1

[0077] Δx 23 =x 3 -x 2

[0078] Δy 23 =y3 -y 2

[0079] Δx 31 =x 1 -x 3

[0080] Δy 31 =y 1 -y 3

[0081] Then the coordinates of the boundary points of the three sides of the triangle are:

[0082] (x 12 ,y 12 )=(x 1 +kΔx 12 ,y 1 +kΔy 12 )

[0083] (x 23 ,y 23 )=(x 2 +kΔx 23 ,y 2 +kΔy 23 )

[0084] (x 31 ,y 31 )=(x 3 +kΔx 31 ,y 3 +kΔy 31 )

[0085] Where k is an integer, which indicates the step length along the edge, usually 1 or -1;

[0086] y min =min(y 1 ,y 2 )

[0087] y max =max(y 1 ,y 2 )

[0088] Among them, y min and max is the ordinate of the lowest and highest points of the triangle, then the coordinates of the colored points inside the triangle are:

[0089] (x,y)=(x left +l(x right -x left ),y)

[0090] Where y is an integer, indicating the vertical coordinate of the scan line, and its range is [y min ,y max ]; x left and x right are the horizontal coordinates of the intersection of the scan line and the three sides of the triangle, which can be obtained by interpolation; l is a decimal, which represents the ratio of the dyed points on the scan line between the left and right intersections, and its range is [0,1]; the principle of the bidirectional scanning algorithm is to start from the lowest point of the triangle and scan line by line along the vertical axis, each time finding the intersection of the scan line and the three sides of the triangle, and then dyeing evenly between the intersections to ensure that all points inside the triangle are dyed; finally, the color value of the pixel area on the screen is reflected to the vertex or fragment of the two-dimensional plane, thereby realizing the update of the texture; in this way, each triangle is assigned a new color value, that is, a color block.

[0091] Furthermore, in step S4, the improved ant colony fusion DWA algorithm is used to solve the shortest path of the roller (paving machine), and the calculation method is:

[0092] Ant k chooses the direction of travel from the "completely compacted" area i to the "partially compacted" or "uncompacted" area j by calculating the state transition probability at time t, and calculates the transition probability of the optional path of the adjacent grid of area i by the formula The transition probability of transition grid region j is determined using the roulette rule as follows:

[0093]

[0094] Among them, η ij (t) is the heuristic function; allow k (k=1, 2, ..., m) is the set of unfinished rolling areas of ant k; α is the pheromone importance factor, referred to as the information factor; β is the heuristic function importance factor, referred to as the heuristic function factor;

[0095] To prevent excessive pheromones from misleading subsequent ants in choosing paths, after each generation of ants completes pathfinding, the information on the path will be updated according to the following formula:

[0096] τ ij (t+1)=(1-ρ)τ ij (t)+Δτ ij (t,t+1)

[0097] Where: ρ is the pheromone volatility coefficient; 1-ρ is the pheromone residual factor; Δτ ij (t, t+1) is the pheromone increment between i and j in this cycle;

[0098] The present invention adopts an improved initial pheromone distribution method, which calculates the size of the initial pheromone according to the number of obstacles around the node, and is specifically implemented as follows:

[0099] τ ij (0) = γ * f(j);

[0100] f(j)=1-C U (allowed j ) / 8

[0101] Where: γ is the enhancement coefficient, which can be set according to the actual situation; f(j) is the obstacle avoidance coefficient; C U is the complement symbol; U is the set of adjacent grids of the current grid; the f(j) function can be used to calculate the proportion of free grids in the 8 adjacent grids of the current grid. When the number of obstacle grids in the neighborhood grid of the grid is smaller, the initial pheromone of the grid is larger; conversely, the initial pheromone is smaller; this guides the ants to move forward, avoids ants from finding paths on terrain with too many obstacles, and speeds up the convergence of the algorithm;

[0102] In the ant colony algorithm, the heuristic information only considers the Euclidean distance between the candidate node and the target point, which will cause the ants to be more attracted in the area close to the target point. However, in the area far away from the target point, the ants cannot be guided by the target point, which will lead to the randomness of the ant path finding in the early stage. Therefore, in order to improve the search heuristics, the Euclidean distance between the current node and the next node is added to the heuristic function. In addition, in order to weaken the influence of heuristic information on path selection in the later stage and speed up the convergence of the algorithm, the improved heuristic function is as follows:

[0103]

[0104] Where: Nc max is the maximum number of iterations, and Nc is the current number of iterations.

[0105] Furthermore, the algorithm flow of using the improved ant colony fusion DWA algorithm to solve the shortest path of the paver (road roller) in step S4 is:

[0106] H1: Initialize relevant parameters;

[0107] H2: Place ants at different starting points at the edge of the construction area, and calculate the next unfinished rolling area that each ant will reach until the updated pheromone table has ants that have visited all unfinished rolling areas;

[0108] H3: Calculate the length of the path that each ant has passed, record the optimal solution in the current iteration, and update the pheromone concentration on the connecting path of each unfinished rolling area;

[0109] H4: Determine whether the maximum number of iterations has been reached. If not, return to step C2, otherwise terminate the program;

[0110] H5: Output the program result, that is, the shortest path, and upload the obtained shortest path to the database.

[0111] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0112] 1. The method of the present invention utilizes drone inspection technology to collect road surface image data in real time, and performs intelligent analysis through deep learning algorithms, providing the construction team with accurate construction quality assessment, effectively improving the monitoring accuracy and response speed of the construction process, and ensuring high standards of construction quality and high efficiency of the construction process.

[0113] 2. The present invention uses a deep learning model to automatically identify defects and evaluate construction quality on images collected by drones, reducing reliance on manual inspections and saving human resources. At the same time, it improves the objectivity and scientific nature of construction monitoring and enhances the adaptability and flexibility of construction plans.

[0114] 3. The present invention combines drone inspections with deep learning algorithms to achieve comprehensive coverage and detail capture of the construction area. It provides construction optimization suggestions to the construction team through an intelligent recommendation system, improves the intelligence level of construction decision-making, optimizes resource allocation, and improves construction efficiency and quality.

[0115] 4. The deep learning model of the present invention has the ability of self-learning and continuous optimization. It can be continuously adjusted and improved according to new data, adapting to changes in different construction environments and conditions, ensuring the continuous updating and optimization of construction plans, and enhancing the long-term applicability and foresight of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 is a flow chart of the method of the present invention;

[0117] Figure 2 This is the network structure after fine-tuning of Faster-RCNN;

[0118] Figure 3 Schematic diagram of the improved ant colony fusion DWA algorithm;

[0119] Figure 4 For raster maps; DETAILED DESCRIPTION

[0120] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0121] The present invention provides an intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, such as Figure 1 As shown, it includes the following steps:

[0122] S1: Collect information such as the center stake coordinates, road width, curve elements, etc. of the highway, draw a two-dimensional plane map of the corresponding section of the highway and generate the corresponding grid map. At the same time, ensure the real-time update of the grid map during the construction process. The specific process is as follows:

[0123] A1: Based on the curve elements and other information collected by S1, the direction and position of the road centerline, as well as the horizontal and vertical curve elements of the road, such as straight lines, circular curves, transition curves, etc., can be determined. The expression formula of the transition curve is as follows:

[0124]

[0125] Among them, s is the arc length on the transition curve, A is the transition curve parameter, and L is the transition curve length. These parameters can be obtained from the curve element;

[0126] A2: Based on the elements of the road centerline, calculate the coordinates and mileage of each point, as well as the tangent angle, curvature radius, curve length and other parameters of the curve to form a stake-by-stake coordinate table. Generally speaking, the content of the stake-by-stake coordinate table includes the following parts: stake number, coordinates, elevation, tangent angle, curvature radius;

[0127] A3: Use professional mapping software such as CASS to draw the plan view and longitudinal section view of the road centerline, as well as the graphics of the road sideline, roadbed, pavement and other parts according to the pile-by-pile coordinate table, convert the numerical data of the road into graphic data, and finally output a two-dimensional plan view;

[0128] A4: The drawn two-dimensional plane map is divided into square cells with a side length of 5 cm. Each cell is called a grid. Each grid represents an area in the construction section and can be used to represent obstacles, drivable space and other information. When the grid map is constructed, each roller (paver) is regarded as an observation point to facilitate the subsequent calculation of a reasonable construction path.

[0129] The specific method to generate a raster map is:

[0130] B1: The observation data of the construction site is obtained through the sensors installed on the roller (paving machine), that is, the geographic coordinates and direction of the roller (paving machine), and the position of each observation point in the corresponding coordinate system and the corresponding grid number are converted according to the geographic coordinates;

[0131] B2: For the observation point, determine whether it is about to touch the obstacle and which grids it has passed along the way based on its distance and angle. For the grid that is about to touch the obstacle, update its status to Occupied, for the grid that has passed, update its status to Free, and for the grid that has not been observed, keep its status as Unknown;

[0132] B3: In order to deal with the noise and uncertainty of the sensor, as well as the dynamic changes of the environment, a probabilistic method is usually used to update the state of the grid;

[0133] B4: By repeating steps B2 to B3, a grid map of the environment can be gradually constructed and updated according to real-time observation data. Finally, it can be determined whether it is an obstacle or free space based on the state of the grid.

[0134] The specific method to update the grid map status is:

[0135] Occupancy Grid Map, which uses the Bayesian rule to calculate the probability of each grid being occupied, is iteratively updated based on observed data and prior knowledge. The update formula of the occupancy grid map is as follows:

[0136]

[0137] Among them, P(m i ∣z 1:t ,x 1:t ) represents the grid state sequence z at all given time steps t 1:t and all observation sequences x at time step t 1:t Under the condition of i The probability of being occupied, z 1:t represents the grid state sequence from time step 1 to t, x 1:t represents the observed data sequence from time step 1 to t;

[0138] In order to simplify the calculation, the log-likelihood ratio is usually used to represent the state of the grid, that is:

[0139]

[0140] In this way, the state of the grid can be represented by a real number instead of a probability value; the larger the state of the grid, the more likely it is to be occupied; a grid state of 0 means that the probability of the grid being idle and occupied is equal or unknown;

[0141] The benefit of using the log-likelihood ratio is that the update formula can be simplified to:

[0142] l i = l i +l occu (z t ,x t )-l 0

[0143] In this way, each time the state of the grid is updated, it is only necessary to add a constant to the current log-likelihood ratio to obtain the latest log-likelihood ratio without considering the previous observation data and the position of the roller (paver); this can greatly improve the computational efficiency and save storage space.

[0144] S2: Use the Road Surface Classification Dataset (RSCD) released by Tsinghua University to train the deep learning model, use indicators such as precision (P), recall (R), F1-Score and mean average precision (mAP) to compare and verify the model, continuously optimize the model parameters, and select the deep learning model with the best effect. The specific process is as follows:

[0145] C1: Preprocess the Road Surface Classification Dataset (RSCD), which is required to include the gradation information of rockfill materials and label it;

[0146] C2: Divide the preprocessed data set into training set, validation set, and test set to ensure that the model can be learned and validated on data at different stages;

[0147] C3: Select a suitable deep learning model, such as Faster-RCNN or YOLO series algorithms, and train the model on the training set, and use the validation set to adjust and optimize the model parameters;

[0148] C4: Use the test set to evaluate the performance of the trained model. Use evaluation indicators such as precision (P), recall (R), F1-Score, and mean average precision (mAP) to comprehensively compare and verify the performance of the model. After comparison, we found that Faster-RCNN has better performance.

[0149] C5: Based on the evaluation results, iteratively optimize the model, which may include adjusting the network structure, improving the training strategy, and introducing regularization methods to improve the accuracy and robustness of the model;

[0150] C6: During the iterative optimization process, the validation set is continuously used to evaluate the model performance to ensure that the model does not overfit to the training data while maintaining high accuracy;

[0151] C7: Faster-RCNN was finally selected as the model for collecting rockfill gradation information in the compaction area, and was validated on the test set to ensure its effectiveness and reliability in practical applications.

[0152] S3: Use a high-resolution camera on a drone to capture images of the construction road surface, and use the best deep learning model to analyze the images. Collect the gradation information of the road surface rockfill and predict the density of the corresponding rolling area based on the fractal model of continuous gradation. Update the color of the corresponding area on the two-dimensional plane according to the density of different areas to visualize the rolling situation of each area. The specific process is as follows:

[0153] D1: Integrate high-precision positioning systems such as the Global Positioning System (GPS) and Inertial Measurement Unit (IMU) on the drone to ensure that the road images captured accurately correspond to the actual geographic location and record the flight path and shooting posture of the drone;

[0154] D2: The high-resolution camera on the drone collects high-definition images of the construction road surface according to the preset flight altitude and speed, ensuring that the image coverage is comprehensive and has sufficient resolution;

[0155] D3: The collected image data is transmitted to a server or cloud platform in real time through a wireless communication module, such as Wi-Fi or Bluetooth, for further processing and analysis;

[0156] D4: The server uses the Faster-RCNN model to extract the gradation information of the rockfill material from the transmitted image data, and then uses the continuous gradation fractal model to process the gradation information of the rockfill material and evaluate the density, thus realizing intelligent evaluation of the pavement compaction quality.

[0157] D5: After collecting the density corresponding to each area, assign corresponding colors to different areas according to the density.

[0158] The specific method for evaluating the compactness in step D4 is:

[0159] E1: According to the rockfill grading information collected by D4 and the fractal theory, the particle size distribution and particle mass distribution of the rockfill can be described by the fractal dimension D. The particle number N is related to the maximum particle size d maxThe relationship between can be expressed as a fractal distribution formula:

[0160]

[0161] Among them, N 0 is a constant, D is the fractal dimension, and d is the diameter of the particle;

[0162] Similarly, the particle mass P and the maximum particle size d max The relationship between can also be expressed as a fractal distribution formula:

[0163]

[0164] Among them, C 0 is a normalization constant that ensures that when d = d max When , P(d)=1;

[0165] In actual engineering, grading has a significant impact on the density of rockfill materials; the fractal dimension D of grading can reflect the uniformity of particle size distribution, which in turn affects the density after compaction; when the fractal dimension D is low, it means that the particle distribution is relatively uneven, there may be large gaps, resulting in lower density; when the fractal dimension D is high, it means that the particle distribution is relatively uniform, the gaps between particles are small, and a higher density can be achieved;

[0166] E2: Studies have shown that there is an optimal range of fractal dimension D within which the rockfill material can achieve maximum density; this range is usually between 2.22 and 2.63; within this range, the particle size distribution can achieve optimal inter-particle filling, thereby obtaining the best compaction effect;

[0167] E3: The density ρ can be expressed by the following formula:

[0168]

[0169] Where m is the mass of the rockfill in dry state, V total is the total volume of the rockfill, V voids is the volume of pores inside the rockfill, P is the porosity, K particle is the particle distribution coefficient, K water is the moisture content coefficient;

[0170] By adjusting the gradation of rockfill materials, which includes the number of particles N and the particle mass P, the fractal dimension D can be indirectly optimized, thereby improving the density; the optimization can be achieved by adjusting the ratio of particles of different particle sizes to achieve the best inter-particle filling effect.

[0171] In step D5, after collecting the density corresponding to each area, the specific method of assigning corresponding colors to different areas according to the density is as follows:

[0172] F1: First, we use image segmentation techniques to separate different regions in the image, which represent different rolling stages or pavement materials;

[0173] F2: Next, we identify each separated area and determine the road surface condition they represent based on the compaction level; this process includes identifying whether the road surface has been compacted, the degree of compaction, and whether there are areas that need further treatment;

[0174] F3: After identification, we will merge regions with similar characteristics. All regions identified as "compacted" can be merged together for clearer representation in the visualization.

[0175] F4: Finally, we merge the merged area with the original image and update the color of the corresponding area on the 2D plane. This step uses a specific algorithm to adjust the color in the image to reflect the compaction condition of each area; the color update can intuitively reflect the compaction quality of the road surface.

[0176] The specific method of updating the color using a specific algorithm in step F4 is:

[0177] G1: The number and level of colors, as well as the corresponding relationship of colors, can be determined according to the different compaction qualities. For example, the compaction quality can be divided into three levels: "uncompacted", "partially compacted" and "fully compacted", corresponding to the colors red, light green and dark green respectively;

[0178] G2: Determine the boundaries of the dyeing area according to the Bresenham algorithm, determine the dyeing points through the bidirectional scanning algorithm, and then dye each area.

[0179] The specific method of determining the boundary of the dyeing area according to the Bresenham algorithm and determining the dyeing point by the bidirectional scanning algorithm is as follows:

[0180] The formulas of the Bresenham algorithm for determining the boundaries of the dyed area and the bidirectional scanning algorithm for determining the dyed points are as follows:

[0181] Assume the screen coordinates of the three vertices of the triangle are (x1, y1), (x2, y2), (x3, y3), then:

[0182] Δx 12 =x 2 -x 1

[0183] Δy 12 =y 2 -y1

[0184] Δx 23 =x 3 -x 2

[0185] Δy 23 =y 3 -y 2

[0186] Δx 31 =x 1 -x 3

[0187] Δy 31 =y 1 -y 3

[0188] Then the coordinates of the boundary points of the three sides of the triangle are:

[0189] (x 12 ,y 12 )=(x 1 +kΔx 12 ,y 1 +kΔy 12 )

[0190] (x 23 ,y 23 )=(x 2 +kΔx 23 ,y 2 +kΔy 23 )

[0191] (x 31 ,y 31 )=(x 3 +kΔx 31 ,y 3 +kΔy 31 )

[0192] Where k is an integer, which indicates the step length along the edge, usually 1 or -1;

[0193] y min =min(y 1 ,y 2 )

[0194] y max =max(y 1 ,y 2 )

[0195] Among them, y min and maxis the ordinate of the lowest and highest points of the triangle, then the coordinates of the colored points inside the triangle are:

[0196] (x,y)=(x left +l(x right -x left ),y)

[0197] Where y is an integer, indicating the vertical coordinate of the scan line, and its range is [y min ,y max ]; x left and x right are the horizontal coordinates of the intersection of the scan line and the three sides of the triangle, which can be obtained by interpolation; l is a decimal, which represents the ratio of the dyed points on the scan line between the left and right intersections, and its range is [0,1]; the principle of the bidirectional scanning algorithm is to start from the lowest point of the triangle and scan line by line along the vertical axis, each time finding the intersection of the scan line and the three sides of the triangle, and then dyeing evenly between the intersections to ensure that all points inside the triangle are dyed; finally, the color value of the pixel area on the screen is reflected to the vertex or fragment of the two-dimensional plane, thereby realizing the update of the texture; in this way, each triangle is assigned a new color value, that is, a color block.

[0198] S4: Based on the density assessment results, determine which areas have insufficient density and need further compaction. If there are areas with insufficient density, calculate the shortest driving path for each roller (paver) to the area that does not meet the density requirement. Plan the driving path of each roller (paver) according to the principle of shortest path and largest coverage of compaction area.

[0199] The improved ant colony fusion DWA algorithm is used to solve the shortest path of the roller (paving machine). The calculation method is:

[0200] Ant k chooses the direction of travel from the "completely compacted" area i to the "partially compacted" or "uncompacted" area j by calculating the state transition probability at time t, and calculates the transition probability of the optional path of the adjacent grid of area i by the formula The transition probability of transition grid region j is determined using the roulette rule as follows:

[0201]

[0202] Among them, η ij (t) is the heuristic function; allow k (k=1, 2, ..., m) is the set of unfinished rolling areas of ant k; α is the pheromone importance factor, referred to as the information factor; β is the heuristic function importance factor, referred to as the heuristic function factor;

[0203] To prevent excessive pheromones from misleading subsequent ants in choosing paths, after each generation of ants completes pathfinding, the information on the path will be updated according to the following formula:

[0204] τ ij (t+1)=(1-ρ)τ ij (t)+Δτ ij (t,t+1)

[0205] Where: ρ is the pheromone volatility coefficient; 1-ρ is the pheromone residual factor; Δτ ij (t, t+1) is the pheromone increment between i and j in this cycle;

[0206] The present invention adopts an improved initial pheromone distribution method, which calculates the size of the initial pheromone according to the number of obstacles around the node, and is specifically implemented as follows:

[0207] τ ij (0) = γ * f(j);

[0208] f(j)=1-C U (allowed j ) / 8

[0209] Where: γ is the enhancement coefficient, which can be set according to the actual situation; f(j) is the obstacle avoidance coefficient; C U is the complement symbol; U is the set of adjacent grids of the current grid; the f(j) function can be used to calculate the proportion of free grids in the 8 adjacent grids of the current grid. When the number of obstacle grids in the neighborhood grid of the grid is smaller, the initial pheromone of the grid is larger; conversely, the initial pheromone is smaller; this guides the ants to move forward, avoids ants from finding paths on terrain with too many obstacles, and speeds up the convergence of the algorithm;

[0210] In the ant colony algorithm, the heuristic information only considers the Euclidean distance between the candidate node and the target point, which will cause the ants to be more attracted in the area close to the target point. However, in the area far away from the target point, the ants cannot be guided by the target point, which will lead to the randomness of the ant path finding in the early stage. Therefore, in order to improve the search heuristics, the Euclidean distance between the current node and the next node is added to the heuristic function. In addition, in order to weaken the influence of heuristic information on path selection in the later stage and speed up the convergence of the algorithm, the improved heuristic function is as follows:

[0211]

[0212] Where: Nc max is the maximum number of iterations, and Nc is the current number of iterations.

[0213] The algorithm flow of using the improved ant colony fusion DWA algorithm to solve the shortest path of the paver (road roller) is as follows:

[0214] H1: Initialize relevant parameters;

[0215] H2: Place ants at different starting points at the edge of the construction area, and calculate the next unfinished rolling area that each ant will reach until the updated pheromone table has ants that have visited all unfinished rolling areas;

[0216] H3: Calculate the length of the path that each ant has passed, record the optimal solution in the current iteration, and update the pheromone concentration on the connecting path of each unfinished rolling area;

[0217] H4: Determine whether the maximum number of iterations has been reached. If not, return to step C2, otherwise terminate the program;

[0218] H5: Output the program result, that is, the shortest path, and upload the obtained shortest path to the database.

[0219] S5: If the density of the area meets the requirements, continue to evaluate and compact the next area. If the density does not meet the requirements, repeat step S4 until the density of all areas meets the construction standards.

[0220] S6: During the entire construction process, the pavement image dataset is updated in real time, and the deep learning model is retrained and optimized based on the new data to improve the accuracy and reliability of the compactness assessment;

[0221] S7: Determine whether the compaction task of the section has been completed. If not, repeat steps S4 to S6. If the compaction task has been completed and the density of all areas meets the requirements, the work is completed, and the evaluation results provide a basis for construction quality control.

[0222] According to the above examples, in order to address the problems existing in the selection and optimization of road construction plans, such as the limitations of manual experience and calculations and the complexity and variability of construction environments and conditions, the present invention proposes to collect key image data through efficient on-site inspections by drones, and use deep learning algorithms for intelligent analysis to provide real-time and accurate construction quality assessments for construction teams, effectively improving the accuracy and response speed of construction monitoring. The self-learning ability of the deep learning model enables it to continuously optimize and adapt to changes in different construction environments and conditions, ensuring the continuous updating and optimization of construction plans. After the density of the corresponding area meets the standard, the optimal or suboptimal driving route is given through the grid map and path planning algorithm, which improves the intelligence and accuracy of road construction and can adapt to changes in construction needs and conditions; at the same time, the method is highly innovative and scalable, and only a small part of the content needs to be modified before it can be applied in other road construction projects.

Claims

1. An intelligent recommendation method for road construction based on drone inspection and deep learning algorithm, characterized in that: The steps include: S1: Collect the coordinates of the center stakes, road width, and curve element information of the highway, draw a two-dimensional plan of the corresponding section of the highway and generate the corresponding grid map, while ensuring the real-time update of the grid map during the construction process; S2: Use the existing road image dataset to train the deep learning model, use the precision P, recall R, F1-Score and mean average precision mAP indicators to compare and verify the model, continuously optimize the model parameters, and select the deep learning model with the best effect; S3: Use the camera carried by the drone to take images of the construction road surface, and use the best deep learning model selected in step S2 to analyze the images, collect the gradation information of the road surface rockfill material, and predict the density of the corresponding rolling area based on the fractal model of continuous gradation. Update the color of the corresponding area on the two-dimensional plane according to the density corresponding to different areas, so as to visualize the rolling conditions of each area; S4: Based on the compaction evaluation results, determine which areas have insufficient compaction and need further compaction. If there are areas with insufficient compaction, calculate the shortest driving path from each working machine to the area that does not meet the compaction requirements. Plan the driving path of each working machine based on the principle of shortest path and maximum coverage of compaction area. S5: If the density of the area meets the requirements, continue to evaluate and compact the next area. If the density does not meet the requirements, repeat step S4 until the density of all areas meets the construction standards. S6: During the entire construction process, the pavement image dataset is updated in real time, and the deep learning model is retrained and optimized based on the new data to improve the accuracy and reliability of the compactness assessment; S7: Determine whether the compaction task of the bid section has been completed. If not, repeat steps S4 to S6. If the compaction task has been completed and the compaction of all areas meets the requirements, the work is completed, and the evaluation results provide a basis for construction quality control. The density evaluation method is: E1: Based on the collected rockfill grading information and fractal theory, the particle size distribution and particle mass distribution of the rockfill are described by the fractal dimension D, and the particle number N is related to the maximum particle size. The relationship between is expressed as the fractal distribution formula: in, is a constant, D is the fractal dimension, and d is the diameter of the particle; Similarly, the particle mass P and the maximum particle size The relationship between is also expressed as the fractal distribution formula: in, is a normalization constant that ensures that = hour, =1; E2: There is an optimal fractal dimension D range within which the rockfill material achieves the maximum density; this range is between 2.22 and 2.63; within this range, the particle size distribution can achieve the best inter-particle filling, thus obtaining the best compaction effect; E3: Density It is expressed by the following formula: in, is the mass of the rockfill material in dry state, is the total volume of the rockfill, is the volume of pores inside the rockfill, is the porosity, is the particle distribution coefficient, is the moisture content coefficient.

2. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 1 is characterized in that: The method for drawing the two-dimensional plane map and the grid map of the corresponding section of the highway in step S1 is: A1: Based on the information of curve elements collected by S1, determine the direction and position of the road centerline, as well as the elements of the horizontal and vertical curves of the road. The expression formula of the transition curve is as follows: Among them, s is the arc length on the transition curve, A is the transition curve parameter, and L is the transition curve length. These parameters can be obtained from the curve element; A2: Based on the elements of the road centerline, calculate the coordinates and mileage of each point, as well as the tangent angle, curvature radius, curve length and other parameters of the curve to form a stake-by-stake coordinate table. Generally speaking, the content of the stake-by-stake coordinate table includes the following parts: stake number, coordinates, elevation, tangent angle, curvature radius; A3: Use mapping software to draw the plan and longitudinal section of the road centerline, as well as the graphics of the road sideline, roadbed, pavement and other parts according to the pile-by-pile coordinate table, convert the numerical data of the road into graphic data, and finally output a two-dimensional plan; A4: The drawn two-dimensional plane map is divided into square cells with a side length of 5 cm. Each cell is called a grid. Each grid represents an area in the construction section and is used to represent obstacles and drivable space information. When the grid map is constructed, each working machine is regarded as an observation point to facilitate the subsequent calculation of a reasonable construction path.

3. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 2 is characterized in that: The method for generating a raster map in step A4 is: B1: Obtain the observation data of the construction site through the sensors installed on the working machine, that is, the geographic coordinates and direction of the working machine, and convert the geographic coordinates into the position of each observation point in the corresponding coordinate system and the corresponding grid number; B2: For the observation point, determine whether it is about to touch the obstacle and which grids it has passed along the way based on its distance and angle. For the grid that is about to touch the obstacle, update its status to occupied, for the grid that has passed, update its status to idle, and for the grid that has not been observed, keep its status as unknown; B3: In order to deal with the noise and uncertainty of the sensor and the dynamic changes of the environment, a probabilistic method is used to update the state of the grid; B4: By repeating steps B2 to B3, a grid map of the environment is gradually constructed and updated according to real-time observation data. Finally, based on the state of the grid, it is determined whether it is an obstacle or free space.

4. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 3 is characterized in that: The specific method for updating the grid in step B3 is: The occupancy grid map uses the Bayesian rule to calculate the probability of each grid being occupied, and it is iteratively updated based on the observed data and prior knowledge. The update formula of the occupancy grid map is as follows: in, Represents the grid state sequence at all given time steps t and the observation sequence of all time steps t Under the condition of The probability of being occupied, represents the grid state sequence from time step 1 to t, represents the observed data sequence from time step 1 to t; In order to simplify the calculation, the log-likelihood ratio is usually used to represent the state of the grid, that is: At this time, the state of the grid is represented by a real number rather than a probability value; the larger the state of the grid, the more likely it is to be occupied; a grid state of 0 means that the probability of the grid being idle and occupied is equal or unknown; Using the log-likelihood ratio, the update formula is simplified to: In this way, each time the state of the grid is updated, it is only necessary to add a constant to the current log-likelihood ratio to obtain the latest log-likelihood ratio without considering the previous observation data and the position of the working machine.

5. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 1 is characterized in that: The specific process of step S2 is: C1: Preprocess the pavement image dataset, which is required to include the rockfill gradation information and be labeled; C2: Divide the preprocessed data set into training set, validation set, and test set to ensure that the model can be learned and validated on data at different stages; C3: Select a suitable deep learning model and train the model on the training set, and use the validation set to adjust and optimize the model parameters; C4: Use the test set to evaluate the performance of the trained model. Use the precision P, recall R, F1-Score, and average precision mAP evaluation indicators to compare and verify the performance of the model, and finally select the Faster-RCNN model; C5: Based on the evaluation results, the Faster-RCNN model is iteratively optimized, including adjusting the network structure, improving the training strategy, and introducing regularization methods to improve the accuracy and robustness of the Faster-RCNN model. C6: During the iterative optimization process, the validation set is continuously used to evaluate the performance of the Faster-RCNN model to ensure that the model does not overfit to the training data while maintaining high accuracy; C7: Faster-RCNN was finally selected as the model for collecting rockfill gradation information in the compaction area, and was validated on the test set to ensure its effectiveness and reliability in practical applications.

6. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 1 is characterized in that: The specific process of step S3 is: D1: Integrate high-precision positioning systems such as global positioning systems and inertial measurement units on drones to ensure that the road images captured accurately correspond to the actual geographic location and record the flight path and shooting posture of the drone; D2: The high-resolution camera on the drone collects high-definition images of the construction road surface according to the preset flight altitude and speed, ensuring that the image coverage is comprehensive and has sufficient resolution; D3: The collected image data is transmitted to a server or cloud platform in real time through a wireless communication module, such as Wi-Fi or Bluetooth, for further processing and analysis; D4: The server uses the Faster-RCNN model to extract the gradation information of the rockfill material from the transmitted image data, and then uses the continuous gradation fractal model to process the gradation information of the rockfill material and evaluate the density, thus realizing intelligent evaluation of the pavement compaction quality. D5: After collecting the density corresponding to each area, assign corresponding colors to different areas according to the density.

7. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 1 is characterized in that: In step S4, the improved ant colony fusion DWA algorithm is used to solve the shortest driving path from the working machine to the area that does not meet the density requirement. The calculation method is: Ant k chooses the direction of travel from fully compacted area i to partially compacted or uncompacted area j by calculating the state transition probability at time t, and calculates the transition probability of the optional path of the adjacent grid of area i by the formula , the transition probability of the transition grid region j is determined using the roulette rule as follows: in, is the heuristic function; (k=1, 2, ..., m) is the set of areas that ant k has not completed rolling; is the pheromone importance factor, referred to as the information factor; is the importance factor of the heuristic function, referred to as the heuristic function factor; To prevent excessive pheromones from misleading subsequent ants in choosing paths, after each generation of ants completes pathfinding, the information on the path will be updated according to the following formula: Where: is the pheromone volatility coefficient; It is the pheromone residual factor; is the pheromone increment between i and j in this cycle; Then, the improved initial pheromone distribution method is used to calculate the size of the initial pheromone according to the number of obstacles around the current grid. The specific implementation is as follows: Where: is the enhancement coefficient, which is determined according to the actual situation; is the obstacle avoidance coefficient; is the complement symbol; U is the set of adjacent grids of the current grid; use The function calculates the proportion of free grids in the 8 adjacent grids of the current grid. When the number of obstacle grids in the neighborhood grid of the grid is smaller, the initial pheromone of the grid is larger; conversely, the initial pheromone is smaller; this guides the ants to move forward, avoids ants from finding paths on terrain with too many obstacles, and speeds up the convergence of the algorithm; The Euclidean distance between the current node and the next node is added to the heuristic function. In addition, in order to weaken the influence of heuristic information on path selection in the later stage and speed up the convergence of the algorithm, the improved heuristic function is as follows: Where: is the maximum number of iterations, is the current iteration number.

8. The intelligent recommendation method for road construction based on drone inspection and deep learning algorithm according to claim 1 or 7, characterized in that: The algorithm flow of using the improved ant colony fusion DWA algorithm to solve the shortest path of the working machine in step S4 is as follows: H1: Initialize relevant parameters; H2: Place ants at different starting points at the edge of the construction area, and calculate the next unfinished rolling area that each ant will reach until the updated pheromone table has ants that have visited all unfinished rolling areas; H3: Calculate the length of the path that each ant has passed, record the optimal solution in the current iteration, and update the pheromone concentration on the connecting path of each unfinished rolling area; H4: Determine whether the maximum number of iterations has been reached. If not, return to step C2, otherwise terminate the program; H5: Output the program result, that is, the shortest path, and upload the obtained shortest path to the database.

Images