Green space perception evaluation system based on visual depth and landscape features

Abstract

The invention relates to the field of urban landscape perception evaluation and intelligent design, in particular to a green space perception evaluation system based on visual depth and landscape features. The system mainly comprises a data acquisition module, a specific element semantic segmentation module, a monocular vision depth estimation module, a data fusion and perception calculation module and an analysis evaluation and visualization output module, and has the core thought that the visual depth and green visibility index and the perception attenuation of the visual depth and vision in a three-dimensional space are taken as core indexes; a system process integrating five steps of data acquisition, semantic segmentation, depth estimation, fusion calculation and perception evaluation is constructed, and the whole process automation from green seeing to green quantification is realized.

Description

Technical Field

[0001] This invention relates to the field of urban landscape perception assessment and intelligent design, specifically a green space perception assessment system based on visual depth and landscape features. Background Technology

[0002] In urban landscape design and greening assessment, traditional indicators such as "green coverage rate" and "green view rate" are mostly calculated based on area ratios, which cannot accurately reflect the differences in green experience perceived by humans. Especially in scenarios such as street greening, community windows offering views of greenery, and therapeutic landscapes, even if the green area meets the standards, people may not perceive the existence of green spaces due to obstructed views, lack of visual depth, or insufficient layering.

[0003] Existing green space assessment methods have the following main problems: Lack of consideration for visual features: Only area is counted, ignoring human visual habits and visual field depth; Lack of spatial hierarchy description: Unable to quantify the green hierarchy in the foreground, middle ground, and background; Lack of automated analysis capabilities: Relies on manual annotation and subjective judgment, which is time-consuming and has low accuracy.

[0004] To address the aforementioned problems, this invention proposes an automated green space perception and assessment system that combines visual depth information with landscape features. From the perspective of human vision, it constructs a quantifiable green space perception index system by fusing image depth estimation with a semantic segmentation model. Summary of the Invention

[0005] In view of the limitations of the existing methods described above, the purpose of this invention is to propose a green space perception and assessment system based on visual depth and landscape features, so as to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions.

[0006] This invention aims to achieve automated analysis and quantitative assessment of the "perceptibility" of urban green spaces through artificial intelligence and computer vision technology.

[0007] A green space perception and assessment system based on visual depth and landscape features includes the following modules: a data acquisition module, a specific element semantic segmentation module, a monocular visual depth estimation module, a data fusion and perception calculation module, and an analysis, evaluation, and visualization output module. The system's main technical path includes four steps: data acquisition, model training, fusion algorithm, and evaluation method. Data acquisition involves obtaining landscape data through the data acquisition module, which is then fed into the model for training. The model includes a specific element semantic segmentation module and a monocular visual depth estimation module. The trained data is processed by a fusion algorithm within the data fusion and perception calculation module. Finally, the data processed by the fusion algorithm is evaluated and output through the analysis, evaluation, and visualization output module.

[0008] Furthermore, in the data acquisition module, a green visibility index is established by acquiring landscape scene images and then by green space perception assessment and index design. The green space perception assessment is defined as an assessment method based on visual perception and using accessible or usable scenes as observation points. The design of indicators needs to combine biological visual characteristics and attenuation properties, and achieve a complete description of spatial pattern relationships by defining green elements within the field of view and integrating the three dimensions of horizontal, vertical and depth. The green visibility index refers to the accessibility and clarity of green vegetation in a person's field of vision. The green visibility index combines visual habits and landscape characteristics, and is designed around visual dispersion to eliminate perceptual bias caused by overly concentrated distribution of elements. The green visibility index ranges from 0 to 1. The higher the value, the higher the density and balance of green element distribution. This value reflects the distribution of green elements on a two-dimensional visual plane. When superimposed with visual depth information that reflects spatial hierarchy, it can further accurately reflect the distribution quality of green elements in three-dimensional space. Visual depth refers to the three-dimensional distance information between objects or spatial surfaces in a scene and the observer. It can be represented by a depth map, where each pixel value represents the actual distance. It is often normalized to a grayscale image of 0-1 for visualization. Perceptual decay is the process by which the visual perception of green elements gradually weakens with distance and occlusion. ; In the formula, Let be the value of green space perception in the visual depth direction, ranging from [0,1], with larger values ​​indicating stronger perception; d represents visual depth, and ; To detect the attenuation coefficient, a value of 0.02 was determined using a reverse calculation method.

[0009] Substituting the limit of perception into the formula Verification showed that when the corresponding spatial depths for foreground, midground, and background were 6m, 28.5m, and 190m, respectively, the P-values ​​were approximately 0.9, 0.5, and 0.001, which is consistent with the actual attenuation pattern in visual perception where foreground is stronger, midground is weaker, and background is fainter.

[0010] Furthermore, the main steps of the green space perception assessment system include: determining camera parameters based on human visual characteristics and collecting data by uniformly distributing points in the accessible space; simultaneously training a semantic segmentation model for specific elements using a deep learning framework to complete pixel-level segmentation of scene data; combining a monocular visual depth estimation model to overlay depth information onto the segmented data; and finally, using a perception attenuation function to weighted evaluate and output a .npy structure file for pixel statistics and visualization analysis, thereby achieving fully automated analysis of the sampled data. The system's main technical approach includes four modules: data acquisition, model training, fusion algorithm, and evaluation method.

[0011] Furthermore, the specific element semantic segmentation module establishes a segmentation model for training. Compared with other mainstream architectures, the DeepLabV3+ framework used has a DCNN decoder that can flexibly adjust the backbone based on ResNet, reducing training costs; the dilated convolution + ASPP pyramid pooling structure can effectively preserve the receptive field and improve the recognition accuracy of green areas and shrubs as much as possible; the skip connections + lightweight decoder can take into account both deep and shallow features, reduce redundant calculations and preserve clear vegetation edges. The pre-training weights and the label classification system of the training set are different from the target classification system of the GVI specific elements in this study. Label remapping needs to be performed before training. By establishing the correspondence between the three and the target classification label, a semantic segmentation mask containing only "0" and "1" is generated. The Cityscapes and MapillaryVistas datasets are used in combination for transfer training, so that the model pays more attention to the elements that need to be identified in the green space scene.

[0012] Furthermore, the monocular vision depth estimation module extracts spatial depth information from a single two-dimensional image, generating a three-dimensional depth map of the target scene without requiring camera parameters or additional sensor data. Using the open-source monocular depth estimation model Depth Pro, it focuses on the scope and scale of the green space scene discussion, and fine-tunes the model for core elements such as trees, shrubs, and low vegetation. By cleaning and extracting the depth values ​​of the target segment, and normalizing them, the target data is obtained. The monocular vision depth estimation module is optimized through a deep neural network to quickly output high-resolution depth information.

[0013] Furthermore, data fusion focuses on the perceptual attenuation law: initial results are filtered by superimposing the green visibility index and visual depth, and then the perceptual attenuation function is used to weight the reduction of values ​​in the horizontal and vertical directions respectively, establishing a final method for calculating the perceptual level P. ; In the formula, where The label is a separator, and the label is either 0 or 1; , , Weighting coefficients are applied to the horizontal, vertical, and depth dimensions; the horizontal attenuation rate is inversely calculated based on the 4:3 ratio of the major to minor axes of the ellipse representing the human visual field. The value is 0.025. The value is 0.033; These are the horizontal and vertical viewing angle differences, respectively. The perceived intensity in the visual depth direction, after data fusion, is used to evaluate a single green space scene, which includes two outputs: 1. Green Space Perception Distribution Map: The horizontal axis of the histogram represents perception intensity, and the vertical axis represents the number of pixels distributed. From left to right, it corresponds to the distribution of green elements in the foreground, middle ground, and background. Ideally, the histogram shows a uniformly decreasing curve with a higher right side and a lower left side, indicating strong perception of green space in the foreground, rich layers in the middle ground, and natural transition in the background. If there are missing or concave areas, it indicates insufficient green elements in the corresponding depth range.

[0014] 2. Green Space Perception Orientation Map: The polar coordinate map uses the visual center as the origin and presents the average perception intensity in eight directions. It intuitively shows the distribution of perceptual elements in different directions. The further away from the origin, the stronger the perception. The average value is the overall perception intensity of the current scene. According to the principle of perspective, distant scenes are mostly concentrated slightly above the horizon line, medium and near scenes are distributed in a mirror image, and near scenes are at the bottom.

[0015] Furthermore, the analysis, evaluation, and visualization output module outputs the green space perception assessment of landscape features to the display screen in a visualized table format through analysis and evaluation.

[0016] The beneficial effects of this invention are as follows: taking visual depth and green visibility index, visual depth and visual perception attenuation in three-dimensional space as core indicators, a system process integrating five steps, namely data acquisition, semantic segmentation, depth estimation, fusion calculation and perception assessment, is constructed to achieve full automation of the process "from seeing green to quantifying green". Attached Figure Description

[0017] The above and other features of the present invention will become more apparent from the detailed description of the embodiments shown in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 A flowchart of a green space perception assessment system based on visual depth and landscape features; Figure 2 This is a green space perception assessment result diagram of a green space perception assessment system based on visual depth and landscape features; Figure 3 This is a typical scenario calculation result diagram of a green space perception assessment system based on visual depth and landscape features; Figure 4 This is a practical illustration of an office park project using a green space perception assessment system based on visual depth and landscape features. Detailed Implementation

[0018] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0019] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0020] like Figure 1 As shown, a green space perception assessment system based on visual depth and landscape features is disclosed. The system includes the following modules: a data acquisition module, a specific element semantic segmentation module, a monocular visual depth estimation module, a data fusion and perception calculation module, and an analysis, evaluation, and visualization output module. The main technical path of the system includes four steps: data acquisition, model training, fusion algorithm, and evaluation method. The data acquisition module obtains landscape data and inputs the landscape data into the model training module. The model includes a specific element semantic segmentation module and a monocular visual depth estimation module. After model training, the trained data is processed by the fusion algorithm. The data processing runs in the data fusion and perception calculation module. Finally, the data calculated by the fusion algorithm is evaluated and output through the analysis, evaluation, and visualization output module.

[0021] Furthermore, in the data acquisition module, a green visibility index is established by acquiring landscape scene images and then by green space perception assessment and index design. The green space perception assessment is defined as an assessment method based on visual perception and using accessible or usable scenes as observation points. The design of indicators needs to combine biological visual characteristics and attenuation properties, and achieve a complete description of spatial pattern relationships by defining green elements within the field of view and integrating the three dimensions of horizontal, vertical and depth. The green visibility index refers to the accessibility and clarity of green vegetation in a person's field of vision. The green visibility index combines visual habits and landscape characteristics, and is designed around visual dispersion to eliminate perceptual bias caused by overly concentrated distribution of elements. The green visibility index ranges from 0 to 1. The higher the value, the higher the density and balance of green element distribution. This value reflects the distribution of green elements on a two-dimensional visual plane. When superimposed with visual depth information that reflects spatial hierarchy, it can further accurately reflect the distribution quality of green elements in three-dimensional space. Visual depth refers to the three-dimensional distance information between objects or spatial surfaces in a scene and the observer. It can be represented by a depth map, where each pixel value represents the actual distance. It is often normalized to a grayscale image of 0-1 for visualization. Perceptual decay is the process by which the visual perception of green elements gradually weakens with distance and occlusion. ; In the formula, Let be the value of green space perception in the visual depth direction, ranging from [0,1], with larger values ​​indicating stronger perception; d represents visual depth, and ; To detect the attenuation coefficient, a value of 0.02 was determined using a reverse calculation method.

[0022] Substituting the limit of perception into the formula Verification showed that when the corresponding spatial depths for foreground, midground, and background were 6m, 28.5m, and 190m, respectively, the P-values ​​were approximately 0.9, 0.5, and 0.001, which is consistent with the actual attenuation pattern in visual perception where foreground is stronger, midground is weaker, and background is fainter.

[0023] Furthermore, the main steps of the green space perception assessment system include: determining camera parameters based on human visual characteristics and collecting data by uniformly distributing points in the accessible space; simultaneously training a semantic segmentation model for specific elements using a deep learning framework to complete pixel-level segmentation of scene data; combining a monocular visual depth estimation model to overlay depth information onto the segmented data; and finally, using a perception attenuation function to weighted evaluate and output a .npy structure file for pixel statistics and visualization analysis, thereby achieving fully automated analysis of the sampled data.

[0024] The system's main technical approach includes four modules: data acquisition, model training, fusion algorithm, and evaluation method.

[0025] Furthermore, the specific element semantic segmentation module establishes a segmentation model for training, using the DeepLabV3+ framework; The pre-trained weights and the label classification system of the training set need to be re-mapped before training. By establishing the correspondence between the three and the target classification label, a semantic segmentation mask containing only "0" and "1" is generated. The Cityscapes and MapillaryVistas datasets are used together for transfer training.

[0026] The pre-training weights and the label classification system of the training set are different from the target classification system of the GVI specific elements in this study. Label remapping needs to be performed before training. By establishing the correspondence between the three and the target classification label, a semantic segmentation mask containing only "0" and "1" is generated. The Cityscapes and MapillaryVistas datasets are used in combination for transfer training, so that the model pays more attention to the elements that need to be identified in the green space scene.

[0027] Furthermore, a 3D depth map of the target scene is generated through the monocular visual depth estimation. Within this 3D depth map, the open-source monocular depth estimation model Depth Pro is used to fine-tune the scale of the focused green space scene within the target scene, as well as the core elements of trees, shrubs, and low vegetation. The depth values ​​of the target segments are cleaned and extracted, and then normalized to obtain the target data. The monocular visual depth estimation module is optimized through a deep neural network to quickly output high-resolution depth information. This approach balances the accuracy and speed limitations of monocular depth estimation, providing lightweight technical support for depth information extraction in the evaluation process.

[0028] Furthermore, data fusion focuses on the perceptual attenuation law: initial results are filtered by superimposing the green visibility index and visual depth, and then the weighted horizontal and vertical values ​​are weakened respectively through the perceptual attenuation function to establish the final perceptuality P calculation method. ; In the formula, where The label is a separator, and the label is either 0 or 1; , , Weighting coefficients are applied to the horizontal, vertical, and depth dimensions; the horizontal attenuation rate is inversely calculated based on the 4:3 ratio of the major to minor axes of the ellipse representing the human visual field. The value is 0.025. The value is 0.033; These are the horizontal and vertical viewing angle differences, respectively. The perceived intensity in the visual depth direction, after data fusion, results in a single green space scene assessment with two outputs. For example... Figure 2 As shown, the two results are: 1. Green Space Perception Distribution Map: The horizontal axis of the histogram represents perception intensity, and the vertical axis represents the number of pixels distributed. From left to right, it corresponds to the distribution of green elements in the foreground, middle ground, and background. Ideally, the histogram shows a uniformly decreasing curve with a higher right side and a lower left side, indicating strong perception of green space in the foreground, rich layers in the middle ground, and natural transition in the background. If there are missing or concave areas, it indicates insufficient green elements in the corresponding depth range.

[0029] 2. Green Space Perception Orientation Map: The polar coordinate map uses the visual center as the origin and presents the average perception intensity in eight directions. It intuitively shows the distribution of perceptual elements in different directions. The further away from the origin, the stronger the perception. The average value is the overall perception intensity of the current scene. According to the principle of perspective, distant scenes are mostly concentrated slightly above the horizon line, medium and near scenes are distributed in a mirror image, and near scenes are at the bottom.

[0030] Furthermore, the analysis, evaluation, and visualization output module outputs the green space perception assessment of landscape features to the display screen in a visualized table format through analysis and evaluation.

[0031] Figure 3 Six typical landscape scenarios were selected for evaluation, verifying the applicability of the system in different landscape scenarios. The characteristics are as follows: Natural type: high level of perception, rich vegetation layers; histogram decreases uniformly, polar coordinate graph shows an outward convex polygon; Hybrid type: It lies between artificial and natural vegetation, with relatively concentrated and layered vegetation; the histogram shows a concentrated distribution; the polar coordinate graph is a semi-circle or a convex polygon with missing corners. Artificial type: high proportion of non-natural elements, green spaces have artificial characteristics; histogram distribution is relatively sharp; polar coordinate graph bulges sharply in one or more directions; Channel type: mainly road paving, with no greenery underneath; histogram drops sharply; polar graph is a horizontal "8" or "C" shape; Symmetrical type: symmetrical greening design, similar left and right layout; histogram is uniform or drops sharply; polar coordinate graph is symmetrical left and right; Covering type: Unnatural elements at the top are obscured; the histogram drops sharply; the polar plot shows a polygon missing at the top.

[0032] Figure 4 In the office park project shown, limited soil cover made it difficult to plant a large number of tall trees, and the concentrated fire-fighting area resulted in a high proportion of hard paving. How to effectively improve the perceived greenness, balance ecological benefits, and enhance the spatial experience through scheme optimization became the core design challenge.

[0033] This paper applies the theories and algorithms presented to construct an evaluation model. All accessible observation points are extracted from the model space, and six-way cameras are deployed in batches to collect visual imaging data. The data is then weighted and integrated into the corresponding collection points for comprehensive perception analysis, thereby adjusting the category, scale, and location of landscape elements. Finally, a three-step implementation strategy is proposed: "prioritizing large trees, supplementing with shallow-rooted plants, and adding movable tree pits." This achieves quantitative analysis of subjective perception and enhances the credibility of the scheme by comparing data before and after the design, forming a closed-loop method for improving green perception under limited conditions.

Claims

1. A green space perception and assessment system based on visual depth and landscape features, characterized in that, The system comprises the following modules: a data acquisition module, a specific element semantic segmentation module, a monocular vision depth estimation module, a data fusion and perception calculation module, and an analysis, evaluation, and visualization output module. The system's main technical path includes four steps: data acquisition, model training, fusion algorithm, and evaluation method. Data acquisition involves obtaining landscape data through the data acquisition module, which then feeds the landscape data into the model for training. The model includes a specific element semantic segmentation module and a monocular vision depth estimation module. After model training, the trained data is processed by a fusion algorithm within the data fusion and perception calculation module. Finally, the data processed by the fusion algorithm is evaluated and output through the analysis, evaluation, and visualization output module.

2. The green space perception and assessment system based on visual depth and landscape features according to claim 1, characterized in that, In the data acquisition module, a green visibility index is established by acquiring landscape scene images and then by green space perception assessment and index design. The green space perception assessment is defined as an assessment method based on visual perception and using accessible or usable scenes as observation points. The design of indicators needs to combine biological visual characteristics and attenuation properties, and achieve a complete description of spatial pattern relationships by defining green elements within the field of view and integrating the three dimensions of horizontal, vertical and depth. The green visibility index refers to the accessibility and clarity of green vegetation in a person's field of vision. The green visibility index combines visual habits and landscape characteristics, and is designed around visual dispersion to eliminate perceptual bias caused by overly concentrated distribution of elements. The green visibility index ranges from 0 to 1. The higher the value, the higher the density and balance of green element distribution. This value reflects the distribution of green elements on a two-dimensional visual plane. When superimposed with visual depth information that reflects spatial hierarchy, it can further accurately reflect the distribution quality of green elements in three-dimensional space. Visual depth refers to the three-dimensional distance information between objects or spatial surfaces in a scene and the observer. It can be represented by a depth map, where each pixel value represents the actual distance. It is often normalized to a grayscale image of 0-1 for visualization. Perceptual decay is the process by which the visual perception of green elements gradually weakens with distance and occlusion. ； In the formula, Let be the value of green space perception in the visual depth direction, ranging from [0,1], with larger values ​​indicating stronger perception; d represents visual depth, and ; To detect the attenuation coefficient, a value of 0.02 was determined using a reverse calculation method.

3. The green space perception and assessment system based on visual depth and landscape features according to claim 2, characterized in that, The main steps of the green space perception assessment system include: determining camera parameters based on human visual characteristics and collecting data by uniformly distributing points in the accessible space; simultaneously training a semantic segmentation model for specific elements using a deep learning framework to complete pixel-level segmentation of scene data; combining a monocular visual depth estimation model to overlay depth information onto the segmented data; and finally, using a perception attenuation function to weighted evaluate and output a .npy structure file for pixel statistics and visualization analysis, thereby achieving fully automated analysis of the sampled data.

4. The green space perception and assessment system based on visual depth and landscape features according to claim 1, characterized in that, The specific element semantic segmentation module establishes a segmentation model for training, using the DeepLabV3+ framework; The pre-trained weights and the label classification system of the training set need to be re-mapped before training. By establishing the correspondence between the three and the target classification label, a semantic segmentation mask containing only "0" and "1" is generated. The Cityscapes and MapillaryVistas datasets are used together for transfer training.

5. A green space perception and assessment system based on visual depth and landscape features according to claim 1, characterized in that, A 3D depth map of the target scene is generated through the monocular vision depth estimation. In the target scene's 3D depth map, the open-source monocular depth estimation model Depth Pro is used to fine-tune the scale of the focused green space scene in the target scene and the core elements of trees, shrubs, and low vegetation in the target scene's 3D depth map. The depth values ​​of the target segment are cleaned and extracted, and then normalized to obtain the target data. The monocular vision depth estimation module is optimized through a deep neural network to quickly output high-resolution depth information.

6. A green space perception and assessment system based on visual depth and landscape features according to claim 1, characterized in that, Data fusion focuses on the perceptual attenuation law: initial results are filtered by superimposing the green visibility index and visual depth, and then the weighted horizontal and vertical values ​​are weakened separately by the perceptual attenuation function to establish the final perceptuality P calculation method. ； In the formula, where The label is a separator, and the label is either 0 or 1; , , Weighting coefficients are applied to the horizontal, vertical, and depth dimensions; the horizontal attenuation rate is inversely calculated based on the 4:3 ratio of the major to minor axes of the ellipse representing the human visual field. The value is 0.

025. The value is 0.033; These are the horizontal and vertical viewing angle differences, respectively. The perceived intensity in the visual depth direction, the evaluation of a single green space scene after data fusion includes two outputs.

Images