Urban green vision rate evaluation method based on LiDAR point cloud

Abstract

The invention provides an urban green vision rate evaluation method based on LiDAR point cloud, and belongs to the technical field of urban planning and environment evaluation. The method depends on accurate urban point cloud data, firstly performs data preprocessing by combining ground filtering and semantic segmentation technologies, and extracts accurate semantic information; then, the point cloud data are used for green vision rate evaluation and analysis, and complete evaluation of urban point cloud ground filtering-semantic segmentation-green vision rate evaluation and analysis is formed. The method not only overcomes the limitation of a traditional evaluation method, provides an accurate, economic and efficient evaluation means, but also provides a new dimension for the quantification of the greening quality by incorporating the visual perception of people into evaluation, and can provide scientific support for urban planning and vegetation management.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban planning and environmental assessment, and particularly relates to a method for evaluating urban green visibility based on LiDAR point clouds. Background Art

[0002] The level of urban greening is a key indicator of the quality of a city's ecological environment, directly impacting residents' well-being and mental health. Green view rate, a key parameter reflecting the relationship between urban greening effectiveness and human visual perception, has garnered widespread attention in recent years. However, traditional assessment methods, such as those based on street view imagery and manual photography, suffer from incomplete data coverage and insufficient accuracy.

[0003] Traditional green view rate assessment methods are mainly based on the analysis of street view image data. Due to its reliability, accessibility and ease of use, such methods are widely used, such as using Google Street View and Baidu Street View to assess green view rate. However, this method has the following problems: (1) Limited viewing angle: This method mainly relies on images taken from the roadway and cannot accurately reflect the visual experience of vegetation on the sidewalk. Because street view cameras are usually installed on vehicles, the shooting angle mainly reflects the visual experience from the roadway; (2) Insufficient data integrity: The shooting interval of street view cameras is large, which makes it difficult to fully cover the human eye's field of view; (3) Environmental interference: The undulations of the road interfere with data quality.

[0004] To address the above issues, some researchers have used manual photography to collect images on sidewalks for green view rate assessment. Although this method can be closer to the pedestrian's perspective, it has the following drawbacks: (1) Time-consuming and labor-intensive: Manual photography requires a lot of time and manpower; (2) Highly subjective: Different photographers' shooting habits and strategies introduce greater uncertainty, affecting the accuracy and consistency of the assessment; (3) Fixed perspective: Manual photography has a fixed perspective, making it difficult to fully cover the pedestrian's visual range. A large number of pictures must be taken to obtain comprehensive data.

[0005] In recent years, with the development of LiDAR technology, green view rate assessment methods based on point cloud data have gradually become a new solution. For example, in the paper "Assessing the visibility of urban greenery using MLSLiDAR data", LiDAR devices capture the three-dimensional structure of a scene through a laser beam and model the scene in the form of a point cloud, which can effectively avoid interference factors in the acquisition of vehicle-borne street view images. Although the green view rate assessment method based on point cloud has improved accuracy, it still has some problems: (1) it does not fully consider the relationship between occlusion and line of sight; (2) it does not consider the impact of distance on visibility; (3) it is computationally inefficient and difficult to apply to large-scale urban environmental assessments.

[0006] Therefore, developing a method that can accurately and efficiently use LiDAR point cloud data to evaluate green view rate is of great significance for urban planning and greening assessment. Summary of the Invention

[0007] In response to the problems existing in the background technology, the purpose of the present invention is to provide a method for evaluating urban green view rate based on LiDAR point cloud. This method relies on accurate urban point cloud data, and first combines ground filtering and semantic segmentation technology to perform data preprocessing to extract accurate semantic information; then uses the point cloud data to evaluate and analyze the green view rate, forming a complete evaluation of "urban point cloud ground filtering → semantic segmentation → green view rate evaluation and analysis". The method of the present invention not only overcomes the limitations of traditional evaluation methods and provides an accurate, economical and efficient evaluation method, but also provides a new dimension for the quantification of greening quality by incorporating human visual perception into the evaluation, and can provide scientific support for urban planning and vegetation management.

[0008] To achieve the above object, the technical solution of the present invention is as follows:

[0009] A method for evaluating urban green view rate based on LiDAR point cloud includes the following steps:

[0010] Step 1. Use LiDAR equipment to scan urban vegetation and obtain raw point cloud data;

[0011] Step 2. Use deep learning methods to perform ground filtering on the collected raw point cloud data, extract the ground point cloud data, and obtain the line of sight vector parallel to the ground;

[0012] Step 3. Perform semantic segmentation on the ground point cloud data obtained in step 2, and divide the point cloud data into vegetation point cloud data and non-vegetation point cloud data;

[0013] Step 4. Determine the field of view parameters of the observation point based on the theory of human field of view;

[0014] Step 5. Convert the 3D point cloud data into a 2D panoramic image;

[0015] Step 6. Introduce the line of sight attenuation model to optimize the generation of field of view images;

[0016] Step 7. Calculate the green viewing rate value based on the visual field image optimized in step 6, and generate the corresponding green viewing rate distribution thematic map.

[0017] Furthermore, the LiDAR equipment in step 1 includes vehicle-mounted LiDAR, backpack LiDAR, handheld LiDAR, etc.

[0018] Furthermore, step 3 semantic segmentation adopts the point cloud semantic segmentation network DA-Net based on unsupervised domain adaptation, which can effectively identify vegetation features and maintain good generalization ability even between different data sources.

[0019] Furthermore, the field of view parameters include the height of the observation point, the horizontal field of view range and the vertical field of view range; among them, the height of the observation point is 1.65 meters, which corresponds to the average height of Chinese people and simulates the perspective of pedestrians; the horizontal field of view range is set to 360° to fully cover all possible field of view of pedestrians; the vertical field of view range is set to 60°, based on the horizontal direction of the human eye, 30° upward and 30° downward, which is in line with the comfortable field of view range of the human eye.

[0020] Furthermore, the specific process of step 5 is:

[0021] Step 5.1. Construct a coordinate system with the observation point as the origin, the line of sight vector as the x-axis, the horizontal direction as the y-axis, and the pitch direction as the z-axis. Calculate the yaw and pitch of each point in the point cloud data, where yaw is the angle of the point from the line of sight, and pitch is defined as the angle of the point from the xoy plane. The calculation formula is as follows:

[0022]

[0023] Where p i is a point in the point cloud data, and its coordinates are The coordinates of the observation point are (x o ,y o ,z o ), For p i The distance to the observation point, the range of yaw is [-π,π], and the range of pitch is [-fov down ,fov up ], fov down is the maximum downward angle of the line of sight (i.e. the maximum value of the pitch and the negative direction of the z-axis), fov up is the maximum upward angle of the line of sight (i.e. the maximum value of the pitch and the positive direction of the z-axis);

[0024] Step 5.2 removes points outside the field of view, and then sets each point p i Mapped to the pixel A(k i ,l i )middle:

[0025]

[0026] P={p i ∈P∣-π≤yaw i ≤π∪fovdown ≤pitch i ≤fov up} (3)

[0027] Where m and n are the resolutions of the final panoramic image, the length of the image is m, the width is n, and fov down is the maximum downward angle of the line of sight, that is, the maximum value of pitch and the negative direction of the z-axis, fov is the vertical field of view, round(·) is the rounding function; ρ is the resolution of the image; finally, k i The value range is [0,m], l i The value range of is [0,n];

[0028] Step 5.3. Set the pixel A(k i ,l i ), assign point p i The attribute value of is shown in formula (4). When there are multiple points corresponding to the same pixel A(k i ,l i ), take the distance d p The smallest point p i The attribute value of pixel A(k i ,l i ) attribute value;

[0029]

[0030] Class represents the category, that is, vegetation point cloud or non-vegetation point cloud.

[0031] Through the above steps, in the obtained panoramic image, each pixel corresponds to a point in the point cloud and contains its semantic information (vegetation or non-vegetation).

[0032] Furthermore, the specific process of step 6 is:

[0033] The formula of the sight attenuation model obtained by quantifying the exponential function in step 6.1 is as follows:

[0034]

[0035] Step 6.2. Optimize the field of view image:

[0036] Calculate the weight of the point in the point cloud. The calculation formula is:

[0037]

[0038] When pixel A(k i ,l i ), assign point p i When the attribute value of the point is reached, the weight value of the point is added to the pixel A(ki ,l i ), as shown in formula (7),

[0039]

[0040] Step 6.3. After calculating all points in the point cloud, all pixels of the panoramic image A are judged: if the weight is greater than the threshold δ, the pixel A(k i ,l i ) is considered vegetation, otherwise it is considered non-vegetation.

[0041] Furthermore, the calculation formula for the green viewing rate in step 7 is:

[0042]

[0043] in, is an indicator function that returns 1 when the condition is met, indicating that the pixel is vegetation; otherwise, it returns 0, indicating that the pixel is non-vegetation; m×n is the total number of pixels in the panoramic image.

[0044] Furthermore, the present invention also provides a method for generating a green view rate distribution thematic map based on the green view rate result. The specific process is as follows:

[0045] Set up multiple observation points on the sidewalk, calculate the green view rate value of each observation point, and then generate the green view rate distribution thematic map. i ,l i ) are rendered green, which means that there is vegetation in this field of view; A(k i ,l i ) is 0, it is rendered in red, which means that there is no vegetation in this field of view;

[0046] The green vision rate is divided according to the proportion of pixels with a value of 1 in the entire field of view image. The green vision rate is 0%-15%, and the green vision feeling is "poor"; the green vision rate is 15%-25%, and the green vision feeling is "slightly green"; the green vision rate is 25%-35%, and the green vision feeling is "relatively green"; the green vision rate is above 35%, and the green vision feeling is "very green".

[0047] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0048] 1. This paper proposes a green view rate assessment method based on LiDAR point clouds. This method utilizes point cloud data with precise semantic information, enabling green view rate assessment to be applied to any location within an urban point cloud scene. It is unaffected by random noise (such as road bumps and photographers' shooting habits), thus achieving more accurate green view rate measurement.

[0049] 2. The present invention proposes a line of sight attenuation model, which makes the point cloud-based green view rate assessment method more robust and more in line with the visual perception characteristics of the human eye, thereby improving the accuracy and rationality of the assessment.

[0050] 3. Experiments comparing our method with traditional green view rate assessment methods based on vehicle-mounted street view imagery were conducted on the large-scale open-source datasets KITTI and Toronto-3D. The experimental results demonstrate that our method is comparable in effectiveness and accuracy to traditional methods, while also offering the advantage of being more consistent with pedestrians' visual perception of vegetation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is the technical roadmap of the evaluation method of the present invention.

[0052] Figure 2 Schematic diagram of converting point cloud to panoramic image.

[0053] Figure 3 Schematic diagram of the human eye's field of view for a single pixel.

[0054] Figure 4 Fits a curve to the line of sight attenuation model function.

[0055] Figure 5 Schematic diagram of point cloud data of three survey areas.

[0056] Figure 6 This figure shows the results of using KP-Attention Net for ground filtering tasks.

[0057] Figure 7 This is a visualization of the classification results of the semantic segmentation task using DA-Net.

[0058] Figure 8 This is the observation point trajectory diagram of the green view rate evaluation algorithm.

[0059] Figure 9 This is a thematic map of green viewing rate distribution in the three experimental measurement areas.

[0060] Figure 10 This is a case study on the spatial distribution characteristics of green view rate.

[0061] Figure 11 A comparison chart of the green view rate assessment results based on point cloud and image. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the implementation methods and drawings.

[0063] A method for evaluating urban green view rate based on LiDAR point cloud. The technical roadmap of the method of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0064] Step 1. Use LiDAR equipment to scan urban vegetation and obtain raw point cloud data. LiDAR equipment includes vehicle-mounted LiDAR, backpack LiDAR, handheld LiDAR, etc. Different data collection methods can be selected as needed to obtain more comprehensive urban environment point cloud data.

[0065] Step 2. Use deep learning methods to perform ground filtering on the collected raw point cloud data, extract the ground point cloud data, and obtain the line of sight vector parallel to the ground;

[0066] The purpose of ground filtering is to extract ground point clouds from the complex distribution of urban features. This paper uses deep learning to perform ground point cloud filtering, specifically using the KP-Attention Net network model. This model combines the KPConv convolution kernel with the self-attention mechanism to extract local features while capturing global relationships, thereby achieving high-precision ground point cloud extraction. The ground point cloud data obtained after filtering can be used to calculate the normal vector, providing a basis for the subsequent determination of the line of sight parallel to the ground.

[0067] Step 3. Based on ground filtering, semantic segmentation is performed on the point cloud data, dividing it into vegetation point cloud data and non-vegetation point cloud data. This method uses a point cloud semantic segmentation network (DA-Net) based on unsupervised domain adaptation. This network can effectively identify vegetation features and maintain good generalization capabilities even across different data sources. After semantic segmentation, the point cloud data is assigned precise semantic labels, providing basic data support for green view rate calculation.

[0068] Step 4. Based on human visual field theory, determine the visual field parameters of the observation point. Visual field parameters include the observation point height, horizontal visual field range, and vertical visual field range. The observation point height is 1.65 meters, corresponding to the average height of the Chinese people and simulating the perspective of pedestrians. The horizontal visual field range is set to 360° to fully cover the entire possible visual field of pedestrians. The vertical visual field range is set to 60°, based on the human eye's horizontal direction, with 30° upward and 30° downward, which is consistent with the human eye's comfortable visual field range.

[0069] These parameter settings make the green view rate assessment more consistent with human visual perception and can more accurately reflect pedestrians' green view experience in the city;

[0070] Step 5. Convert the 3D point cloud data into a 2D panoramic image. The schematic diagram of converting point cloud to panoramic image is as follows: Figure 2 The specific method is as follows:

[0071] Step 5.1. With the observation point as the coordinate origin and the x-axis as the line of sight vector, calculate the yaw and pitch of each point in the point cloud data. Yaw is defined as the angle of the point from the line of sight, and pitch is defined as the angle of the point from the xoy plane. The calculation formula is as follows:

[0072]

[0073] Where p i is a point in the point cloud, and its coordinates are The coordinates of the observation point are (x o ,y o ,z o ), For p i The distance to the observation point, the range of yaw is [-π,π], and the range of pitch is [-fov down ,fov up ], fov down is the maximum downward angle of the line of sight (i.e. the maximum value of the pitch and the negative direction of the z-axis), fov up is the maximum upward angle of the line of sight (i.e. the maximum value of the pitch and the positive direction of the z-axis);

[0074] Step 5.2 removes points outside the field of view, and then sets each point p i Mapped to the pixel A(k i ,l i )middle:

[0075]

[0076] Where m and n are the resolutions of the final panoramic image (the length of the image is m and the width is n, as shown in Figure 2 ), fov down is the maximum downward angle of the line of sight (i.e. the maximum value of pitch and the negative direction of the z-axis), fov is the vertical field of view, and round(·) is the rounding function. Finally, k i The value range is [0,m], l i The value range is [0,n].

[0077] Step 5.3. Set the pixel A(k i ,l i ), assign point p i The attribute value of is shown in formula (3). When there are multiple points corresponding to the same pixel A(k i ,l i ), take the distance d p The smallest point p i Attribute value of .

[0078]

[0079] Through the above steps, each pixel in the obtained panoramic image corresponds to a point in the point cloud and contains its semantic information (vegetation or non-vegetation), such as Figure 2 As shown in the image;

[0080] Step 6. Introduce the line of sight attenuation model to optimize the generation of field of view images;

[0081] In the algorithm of step 5, the attributes of each pixel after the point cloud is converted to a panoramic image are determined by the point closest to the observation point and are completely unrelated to points farther away. This will cause two obvious problems in the algorithm:

[0082] (1) The algorithm is greatly affected by noise: if the ambient noise near the observation point is not filtered, then these noise points will cover the real point cloud data far away;

[0083] (2) The attributes of a pixel should be determined by all points within the field of view corresponding to the pixel, rather than the nearest point. Figure 3 As shown, Figure 3 This is a diagram of the human eye's field of view for a single pixel. In this diagram, the human visual perception should be vegetation, but because the point closest to the observation point is non-vegetation, the algorithm in the previous step classifies this pixel as a non-vegetation pixel, which is unreasonable. Figure 3 In the area covered by the orange shadow, the number of vegetation point clouds is obviously greater than that of non-vegetation point clouds, and people's subjective perception should also be that this space is greener.

[0084] Therefore, in order to more accurately simulate the visual perception of the human eye, this paper introduces a line of sight attenuation model to optimize the generation process of panoramic field of view images:

[0085] Step 6.1 Principle of constructing the sight attenuation model: In spatial vision research, the visual influence of an object in the field of view usually decreases gradually with the increase of observation distance. The present invention uses an exponential function to fit the visibility attenuation model, such as Figure 4 As shown, the line of sight attenuation model formula obtained by quantifying the exponential function is as follows:

[0086]

[0087] The correlation coefficient of the fitted curve is 0.9897, indicating the correctness of the line of sight attenuation model of the present invention;

[0088] Step 6.2. In the algorithm of converting the point cloud to the panoramic view image in step 5, the line of sight attenuation model needs to be incorporated: when converting the pixel A(k i ,l i ), assign point p i When the attribute value is , the weight of the point is calculated by formula (5) Then, the weight is added to the pixel A(k i ,l i ), as shown in formula (6);

[0089]

[0090] Step 6.3. After calculating all points in the point cloud, all pixels of the panoramic image A are judged: if the weight is greater than the threshold δ, the pixel A(k i ,l i ) is determined to be vegetation, otherwise it is non-vegetation;

[0091] Through the line of sight attenuation model, the present invention makes nearby points contribute more to pixel attributes, while distant points contribute less, which is consistent with the perception characteristics of human vision and also improves the algorithm's robustness to environmental noise.

[0092] Step 7. Calculate the green view rate value based on the visual field image optimized in step 6 and generate the corresponding green view rate distribution map; the green view rate is defined as the proportion of green elements (vegetation) in the human eye's visual field, and the calculation formula is as follows:

[0093]

[0094] in, is an indicator function that returns 1 when the condition is met and 0 otherwise; m×n is the total number of pixels in the panoramic image;

[0095] Green View Ratio Distribution Map Generation: In practical applications, multiple observation points can be set up on a sidewalk, and the green view ratio value at each observation point can be calculated to generate a green view ratio distribution map. This method divides the green view ratio values ​​into four intervals: 0%-15%: "Poor" green view perception; 15%-25%: "Slightly Green" green view perception; 25%-35%: "Relatively Green" green view perception; and 35% or above: "Very Green" green view perception. This green view ratio distribution map provides an intuitive understanding of the spatial distribution of urban greening, providing a scientific basis for urban planning and greening promotion.

[0096] Example 1

[0097] A method for evaluating urban green view rate based on LiDAR point cloud includes the following steps:

[0098] Step 1: Acquisition of original point cloud data: This example selects three typical areas (the first measurement area, the second measurement area, and the third measurement area) in a certain region for green view rate assessment:

[0099] Step 1.1. Equipment Installation: The LiDAR device is fixed on the roof of the collection vehicle and equipped with a speed sensor, point cloud storage device, and a static differential base station deployed on the ground. Data is collected using a 128-line iScan-SZ LiDAR. The LiDAR is fixed on the roof of the collection vehicle, and the supporting equipment includes a speed sensor, point cloud storage device, and a static differential base station deployed on the ground.

[0100] Step 1.2. Data Collection: The experimental vehicle drives along the campus roads, while recording raw point cloud data, GNSS (Global Navigation Satellite System) data, odometer data, and IMU (Inertial Measurement Unit) data;

[0101] Step 1.3. Data preprocessing: Use StaticToRinex64 software to convert the original RT27 format GNSS data to generate an o file. Convert the o file to gpb format to ensure data compatibility. Combine the ground control point coordinates, vehicle equipment installation parameters, differential positioning base station information, and POS (Position and Orientation System) data for comprehensive calculation.

[0102] Step 1.4. Trajectory calculation: Using GNSS observation data, IMU measurement information, and odometer data, perform inertial-exterior orientation calculation using IE (Inertial Explorer) software to obtain a high-precision trajectory of the captured vehicle and generate a POS file.

[0103] Step 1.5. Static point cloud generation: Input the solved POS file and the original point cloud data into the mmsconvert software, adjust the relevant parameters, and generate complete static 3D point cloud data for the three measurement areas;

[0104] The final point cloud data of the three survey areas are as follows: Figure 5 As shown, it contains complete scene information such as buildings, roads, and vegetation on campus, providing basic data for subsequent green view rate evaluation;

[0105] Step 2: Ground filtering: To obtain the sight vector parallel to the ground, the complete static 3D point cloud data needs to be ground filtered. This embodiment uses KP-Attention Net for ground filtering. The specific implementation process is as follows:

[0106] Step 2.1. Data input: Input the point cloud data of the three survey areas into the KP-Attention Net model respectively;

[0107] Step 2.2. Parameter setting: Set the number of kernel points of the KPConv convolution kernel to 15, the downsampling parameter to 0.06m, the convolution radius to 2.5m, the momentum to 0.98, the initial learning rate to 0.01, and the learning rate to be divided by 10 after every 100 epochs;

[0108] Step 2.3. Filtering: The model extracts local and global features of the point cloud by combining the KPConv convolution kernel with the self-attention mechanism, achieving effective classification of ground points and non-ground points.

[0109] Step 2.4. Calculate the ground normal vector: Use Cloud Compare software to calculate the normal vector of the extracted ground point cloud to provide a basis for the subsequent determination of the line of sight parallel to the ground.

[0110] The filtering results are as follows Figure 6 As shown in the figure, the green point cloud is the extracted ground point cloud; it can be seen from the figure that KP-Attention Net can accurately distinguish between ground and non-ground point clouds, and the ground point cloud is correctly extracted without obvious misjudgment;

[0111] Step 3: Semantic segmentation. Based on ground filtering, semantic segmentation is required for the point cloud data to classify the point cloud into vegetation and non-vegetation. This embodiment uses DA-Net for semantic segmentation. The specific implementation process is as follows:

[0112] Step 3.1. Data Mapping: Map the point cloud data of the three survey areas to the same semantic labeling system as the source domain dataset (SynLiDAR dataset) to ensure the adaptability of the model;

[0113] Step 3.2. Model initialization: Use the DA-Net model pre-trained and converged in the source domain as the teacher network and initialize the student network with the same parameters;

[0114] Step 3.3. Domain Adaptation Training: Perform the UDA (Unsupervised Domain Adaptation) task on the source and target domains (the point cloud data of the three survey areas) to adapt the model to the data distribution characteristics of the target domain.

[0115] Step 3.4. Semantic segmentation execution: semantic segmentation is performed on the point cloud data of the three survey areas to extract vegetation point clouds and non-vegetation point clouds; the prediction results of vegetation categories after semantic segmentation are as follows Figure 7 As shown, different colors represent different semantic categories;

[0116] The quantitative evaluation results are shown in Table 1 below. The average performance indicators of DA-Net in the three measurement areas are as follows: the precision of the vegetation category is 92.82%, the recall rate is 99.69%, and the IoU is 91.57%; the precision of the non-vegetation category is 99.33%, the recall rate is 82.88%, and the IoU is 81.44%; the overall mIoU is 86.51% and the OA is 94.66%. These indicators show that the semantic segmentation method adopted by the present invention can effectively identify vegetation features and provide accurate semantic information for green view rate assessment.

[0117] Table 1 Ablation experiment of self-attention module in KP-Attention Net, unit: %

[0118]

[0119]

[0120] Step 4: Green View Rate Assessment and Thematic Map Generation Based on LiDAR Point Cloud. After acquiring point cloud data with precise semantic information, this embodiment sets up observation points on the sidewalks of the three measurement areas to conduct green view rate assessment. The specific implementation process is as follows:

[0121] Step 4.1. Observation point setting. Set an observation point every 0.5m on the sidewalk. The height of the observation point is set to 1.65 meters (the average height of Chinese people). The observation point position is as follows: Figure 8 As shown, the red track represents the distribution of observation points.

[0122] Step 4.2. Green view rate calculation parameter settings. Horizontal field of view: 360°; Vertical field of view: 60°; Panoramic image resolution: 720×120; Line of sight attenuation model parameters:

[0123] weight function(x) =1.0616×10 -0.0003·x (14)

[0124] Step 4.3. Green Viewing Rate Calculation: For each observation point, the green viewing rate evaluation method of the present invention is applied to calculate its green viewing rate value.

[0125] Step 4.4. Green Vision Ratio Classification: Based on the calculation results, the green vision ratio values ​​are divided into four intervals:

[0126] 0%-15%, the green vision feeling is "poor"; the green vision rate is 15%-25%, the green vision feeling is "slightly green"; the green vision rate is 25%-35%, the green vision feeling is "relatively green"; the green vision rate is above 35%, the green vision feeling is "very green".

[0127] Step 4.5. Generate the green view rate thematic map. Based on the green view rate values ​​of the observation points, generate the green view rate spatial distribution thematic map of each measurement area, such as Figure 9 As shown, the red part of the trajectory represents the green vision rate GVI difference;

[0128] Step 5: Analysis of Green View Ratio Assessment Results. Based on the above assessment results, statistical analysis and visualization of the green view ratios of the three measurement areas are performed to intuitively present the spatial distribution characteristics of the green view ratios:

[0129] Step 5.1. Statistical analysis of green view rate. First measurement area: Approximately 54% of observation points had a green view rate greater than 25%, indicating a good greening level. Second measurement area: Approximately 57% of observation points had a green view rate greater than 25% (relatively green or very green), making pedestrians feel comfortable. Third measurement area: Approximately 40% of observation points had a green view rate less than 15% (poor), indicating a poor greening level.

[0130] Step 5.2. Analysis of spatial distribution characteristics of green view rate. Near building area: Since one side of the field of view is occupied by buildings, the green view rate is usually low, such as Figure 10 (c) As shown in the sidewalk area far away from the building: there are green elements on both sides of the field of view, and the green view rate value is high, such as Figure 10 (a) As shown in the third survey area, the ring belt area: there are buildings on one side, but there is a lack of nearby vegetation on the other side, and the green view rate is low, such as Figure 10 (b)

[0131] Improvement suggestions: For the third survey area, especially its fan-shaped green space, it is recommended to add tall trees or shrubs to increase the green view rate of the area; in areas close to buildings, the proportion of green elements in the field of vision can be improved by adding vertical greening or roof gardens; trees and shrubs can be appropriately added on both sides of the road to create a green landscape with rich layers and enhance the green vision experience of pedestrians.

[0132] The above analysis proves that the LiDAR-based urban green view rate assessment method of the present invention can effectively identify the spatial differences in urban greening levels and provide a scientific basis for urban greening planning and improvement.

[0133] To ensure the scientificity and reliability of the method of the present invention, the present invention verifies the effectiveness of the point cloud-based green view rate assessment method of the present invention from both quantitative and qualitative aspects by comparing it with the traditional image-based green view rate assessment method.

[0134] Given the extensive validation of traditional image green view ratio assessment methods, this study used their results as reliable ground truth. Using the KITTI dataset, a point cloud and image dataset collected simultaneously (at the same time and location), the point cloud data was cropped to a field of view consistent with the image. DeepLab-v3 was used to perform semantic segmentation on the image data to extract vegetation areas. The proposed method was then applied to the point cloud data to evaluate green view ratio.

[0135] The Pearson correlation coefficient (r) was calculated to evaluate the correlation between the two methods; the mean absolute error (MAE) and root mean square error (RMSE) were calculated to quantify the difference between the two methods; the consistency correlation coefficient (CCC) and intraclass correlation coefficient (ICC) were calculated to evaluate the consistency of the two methods; and the systematic deviation of the two methods was evaluated by Bland-Altman analysis. In this way, the measurement effects of the two methods were comprehensively evaluated to ensure the accuracy and reliability of the new method. Several groups of data from multiple experimental scenes were selected to perform calculations and analysis of the evaluation indicators in the previous step to verify the effectiveness of the point cloud-based green view rate evaluation method. Finally, the results of the point cloud-based green view rate evaluation method and the image-based green view rate evaluation method were compared. Figure 11 The verification result statistics are shown in Table 2 below:

[0136] Table 2 Statistical analysis of green vision rates obtained by two methods

[0137]

[0138] Analysis of the tabular data reveals that the mean MAE / RMSE values ​​for each scenario are 0.04 / 0.06, respectively, indicating that the point cloud evaluation method has a low error with the actual value and high accuracy. Both the intraclass correlation coefficient (ICC) (≥0.8) and the consistency correlation coefficient (CCC) (≥0.9) exceed the highest standards, verifying the high consistency and reliability of the evaluation results of the two methods. The correlation coefficients for all scenarios are all >0.94 (particularly reaching 0.96 for the KITTI-00 scenario), indicating a strong correlation between the evaluation results of the different methods. This demonstrates that both methods accurately reflect the green view rate in this scenario.

[0139] The above description is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for evaluating urban green view rate based on LiDAR point cloud, characterized in that: The following steps are involved: Step 1. Use LiDAR equipment to scan urban vegetation and obtain raw point cloud data; Step 2. Use deep learning methods to perform ground filtering on the collected raw point cloud data, extract the ground point cloud data, and obtain the line of sight vector parallel to the ground; Step 3. Perform semantic segmentation on the ground point cloud data obtained in step 2, and divide the point cloud data into vegetation point cloud data and non-vegetation point cloud data; Step 4. Determine the field of view parameters of the observation point based on the theory of human field of view; Step 5. Convert the 3D point cloud data into a 2D panoramic image; Step 6. Introduce the line of sight attenuation model to optimize the generation of field of view images; Step 7. Calculate the green viewing rate value based on the visual field image optimized in step 6, and generate the corresponding green viewing rate distribution thematic map.

2. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The LiDAR device in step 1 includes a vehicle-mounted LiDAR, a backpack-mounted LiDAR, or a handheld LiDAR.

3. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: In step 3, semantic segmentation is performed using the point cloud semantic segmentation network DA-Net based on unsupervised domain adaptation. This network can effectively identify vegetation features and maintain good generalization capabilities even between different data sources.

4. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The field of view parameters include the height of the observation point, the horizontal field of view range, and the vertical field of view range. The height of the observation point is 1.65 meters, which corresponds to the average height of Chinese people and simulates the perspective of pedestrians. The horizontal field of view range is set to 360° to fully cover all possible field of view of pedestrians. The vertical field of view range is set to 60°, based on the horizontal direction of the human eye, 30° upward and 30° downward, which is in line with the comfortable field of view of the human eye.

5. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The specific process of step 5 is: Step 5.

1. Construct a coordinate system with the observation point as the origin, the line of sight vector as the x-axis, the horizontal direction as the y-axis, and the pitch direction as the z-axis. Calculate the yaw and pitch of each point in the point cloud data, where yaw is the angle of the point from the line of sight, and pitch is defined as the angle of the point from the xoy plane. The calculation formula is as follows: Where p i is a point in the point cloud data, and its coordinates are The coordinates of the observation point are (x o ,y o ,z o ), For p i The distance to the observation point, the range of yaw is [-π,π], and the range of pitch is [-fov down ,fov up ], fov down is the maximum downward angle of the line of sight, fov up The maximum upward angle of the line of sight; Step 5.2 removes points outside the field of view, and then sets each point p i Mapped to the pixel A(k i ,l i )middle: P={p i ∈P∣-π≤yaw i ≤π∪fov down ≤pitch i ≤fov up } (3) Where m and n are the resolutions of the final panoramic image, the length of the image is m, the width is n, and fov down is the maximum downward angle of the line of sight, that is, the maximum value of pitch and the negative direction of the z-axis, fov is the vertical field of view, round(·) is the rounding function; ρ is the resolution of the image; finally, k i The value range is [0,m], l i The value range of is [0,n]; Step 5.

3. Set the pixel A(k i ,l i ), assign point p i The attribute value of is shown in formula (4). When there are multiple points corresponding to the same pixel A(k i ,l i ), take the distance d p The smallest point p i The attribute value of pixel A(k i ,l i ) attribute value; Class represents the category, that is, vegetation point cloud or non-vegetation point cloud. Through the above steps, in the obtained panoramic image, each pixel corresponds to a point in the point cloud and contains its semantic information.

6. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The specific process of step 6 is: The formula of the sight attenuation model obtained by quantifying the exponential function in step 6.1 is as follows: Step 6.

2. Optimize the field of view image: Calculate the weight of the point in the point cloud. The calculation formula is: When pixel A(k i ,l i ), assign point p i When the attribute value of the point is reached, the weight value of the point is added to the pixel A(k i ,l i ), as shown in formula (7), Step 6.

3. After calculating all points in the point cloud, all pixels of the panoramic image A are judged: if the weight is greater than the threshold δ, the pixel A(k i ,l i ) is considered vegetation, otherwise it is considered non-vegetation.

7. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The calculation formula for the green viewing rate in step 7 is: in, is an indicator function that returns 1 when the condition is met, indicating that the pixel is vegetation; otherwise, it returns 0, indicating that the pixel is non-vegetation; m×n is the total number of pixels in the panoramic image.

8. The urban green view rate evaluation method based on LiDAR point cloud according to claim 1, characterized in that: The specific generation process of the green view rate distribution thematic map is as follows: Set up multiple observation points on the sidewalk, calculate the green view rate value of each observation point, and then generate the green view rate distribution thematic map. i ,l i ) are rendered green, which means that there is vegetation in this field of view; A(k i ,l i ) is 0, it is rendered in red, which means that there is no vegetation in this field of view; The green vision rate is divided according to the proportion of pixels with a value of 1 in the entire field of view image. The green vision rate is 0%-15%, and the green vision feeling is "poor"; the green vision rate is 15%-25%, and the green vision feeling is "slightly green"; the green vision rate is 25%-35%, and the green vision feeling is "relatively green"; the green vision rate is above 35%, and the green vision feeling is "very green".

Images