A method, device and medium for evaluating public space quality based on multi-dimensional features

Abstract

The present invention discloses a method, device and medium for evaluating the quality of public space based on multidimensional features. The method comprises the following steps: obtaining panoramic images of different types of public spaces in a target area to form a panoramic image data set; performing semantic segmentation on each panoramic image in the panoramic image data set, and processing the data to obtain a semantic segmentation processing result; calculating the global visual, local visual and global spatial feature indicators of each public space according to the semantic segmentation processing result, and generating a quality evaluation index data set for each public space; performing a public space functional facility category detection, and evaluating the function of the public space according to the detection result; and clustering the quality evaluation index data set of each public space to obtain the type of public space. The present invention can quickly and accurately realize the quality evaluation of public space by comprehensively evaluating the quality of public space from multiple dimensions of quality and function.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent measurement of urban public spaces, and in particular to a method, device and medium for evaluating the quality of public spaces based on multidimensional features. Background Art

[0002] Public space refers to urban space used by the public, such as streets, squares, parks, etc. As the core element of urban space organization, urban public space is an important spatial carrier for people's lives and plays an important role in enhancing urban vitality and optimizing urban functional quality. For the evaluation of public space quality, traditional evaluation methods mainly rely on manual field research, which is not only inefficient and costly, but also difficult to achieve large-scale automated evaluation. The results are subjective and lack a standardized quantitative indicator system, making it difficult to make horizontal comparisons.

[0003] Intelligent quality assessment of public space is to identify and quantify landscape elements of public space landscape basic data through machine learning and other methods by using open data collection tools, sensors, image analysis and other technologies, and to achieve objective and automated assessment of public space landscape quality by combining text semantic mining technology, which can effectively improve the efficiency, accuracy and reliability of assessment. However, in the existing public space intelligent quality assessment methods, a single spatial visual feature (such as sky view rate, color spectrum, green view rate, street tortuosity, etc.) is usually used for analysis. A single spatial visual feature is difficult to accurately and comprehensively characterize the quality status of public space. The accuracy and reliability of the actual assessment results still need to be improved, and since the assessment results are usually simple quality grading such as high or low or numerical values, the category of public space cannot be effectively quantified. If the number of spatial visual features is increased to improve the assessment performance, the actual improvement effect is limited and the assessment complexity will be increased, affecting the assessment efficiency and making it difficult to meet the needs of batch assessment. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a public space quality assessment method, equipment and medium based on multi-dimensional features, which can quickly and accurately achieve public space quality assessment by comprehensively assessing the quality of public space from the dimensions of global vision, global space and local vision, and combining the quality and function dimensions.

[0005] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0006] A method for evaluating the quality of public space based on multi-dimensional features, comprising the following steps:

[0007] Acquire panoramic images of different types of public spaces in a target area to form a panoramic image dataset;

[0008] Performing semantic segmentation on each panoramic image in the acquired panoramic image data set using a semantic segmentation model to obtain a semantic segmentation result, and processing the semantic segmentation result to obtain a semantic segmentation processing result, wherein the semantic segmentation processing result includes an area ratio of each type of semantic segmentation element;

[0009] According to the semantic segmentation processing results, the global visual feature index, the local visual feature index and the global spatial feature index of each public space are respectively calculated to evaluate the quality of each public space from the global visual, local visual and global spatial dimensions, and generate a quality evaluation index data set for each public space;

[0010] The global visual feature index is a feature of the overall visual perception of the public space, the local visual feature index is a feature of the specific entity in the perception process of landscape elements, and the global spatial feature index is a feature of the scale and spatial form of the public space;

[0011] Using a deep learning model to detect the categories of public space functional facilities on the semantic segmentation processing results, and evaluating the functions of the public space according to the detection results;

[0012] The quality assessment indicator data set is clustered to implement spatial cluster analysis, and the public space type of each public space is obtained.

[0013] Further, the performing semantic segmentation on each panoramic image in the acquired panoramic image data set using a semantic segmentation model to obtain a semantic segmentation result, and processing the semantic segmentation result to obtain a semantic segmentation processing result includes:

[0014] The semantic segmentation result is obtained by inputting the acquired panoramic image dataset into the semantic segmentation model;

[0015] The semantic segmentation result is used as a segmentation image, and the segmentation image is projected and corrected by using a Hammer-Aitoff equal-product projection transformation method. The areas of various semantic segmentation elements in the segmentation image are calculated according to the projection correction result.

[0016] Furthermore, the use of the Hammer-Aitoff equal-product projection transformation method to perform projection correction on the segmented image includes:

[0017] Establishing a grid system of standardized longitude and latitude coordinates for panoramic images;

[0018] The standardized longitude and latitude coordinates of the established panoramic image are converted into the Hammer coordinate system to convert them into the x and y coordinates of the Hammer-Aitoff projection plane coordinate system. The calculation formula is as follows:

[0019]

[0020] Among them, latitude represents the standardized latitude of the panoramic image, longtitude represents the standardized longitude of the panoramic image, and z is the intermediate calculation quantity used to normalize x and y;

[0021] Reprojecting the segmented image: After separating the RGB channels of the input segmented image, performing the Hammer coordinate system transformation on each color channel independently, and combining the transformed results of the three color channels to generate a projected segmented image;

[0022] A corresponding category index is pre-assigned to each semantic segmentation element category, and a bidirectional mapping relationship between RGB color and category index is constructed, the projected segmented image is read, and the projected segmented image in RGB color space is converted pixel by pixel into the corresponding category index according to the pre-constructed bidirectional mapping relationship between RGB color and category index, so as to obtain a category index map for calculating the area of ​​each type of semantic segmentation element.

[0023] Furthermore, the global visual feature index includes any one or more of a sky view factor, a building view factor, a green view factor, a visual entropy, a skyline fractal dimension (SFD) and color properties. The sky view factor uses the sky pixel ratio to measure the urban building density and sense of enclosure. The building view factor uses the building pixel ratio to measure the sense of enclosure and building density of the urban space. The green view factor uses the green vegetation pixel ratio to measure the degree of greening. The visual entropy is used to measure the complexity of information in the image. The skyline fractal dimension is used to characterize the spatial complexity of the building outline. The color properties include color brightness, contrast and saturation.

[0024] The local visual feature index includes any one or more of the proportion of natural elements, the proportion of traffic elements, the proportion of public facilities, the proportion of safety facilities, the proportion of human and vehicle activities, the proportion of street walls, the proportion of building signs, the proportion of street-facing windows, and the proportion of specified elements. The proportion of natural elements is used to characterize the proportion of natural elements in the image, the proportion of human and vehicle activities is used to characterize the proportion of the number of people and vehicles in the image, and the proportion of specified elements is the proportion of elements in the segmented image result that is greater than a preset proportion threshold or the proportion of elements of a specified type;

[0025] The global spatial feature indicators include any one or more of the proportion of vertical elements, walking width, block length ratio, H / W ratio of street canyons, spatial connectivity and depth map spatial ratio. The vertical element proportion is used to characterize the relationship between building height and block scale, the walking width is the width of the street, the block length ratio is the ratio of block length and width, the H / W ratio of street canyons is the ratio of building height to street width, the spatial connectivity is used to characterize the traffic and connectivity between spaces, and the depth map spatial ratio is the ratio of different space types in the depth map. The panoramic depth map is generated by estimating the depth of the panoramic image using the unlabeled self-distillation v2 model, and the depth map spatial ratio is calculated.

[0026] Furthermore, the public space functional facility category detection based on the semantic segmentation processing result includes: establishing a facility text prompt set, inputting the semantic segmentation processing result into the Grounding DINO model for target detection and positioning; drawing a facility bounding box on the original panoramic image and marking the category label, and outputting the detection result with the marked information, wherein the marked information includes any one or more of the facility category, bounding box coordinates, detection confidence and location information.

[0027] Furthermore, a K-means clustering optimization algorithm based on an adaptive threshold is used to cluster the quality assessment index data set, and the public space types include mixed public spaces, open public spaces and dense commercial blocks, wherein the differences between the various characteristic indicators of the mixed public spaces and the various characteristic indicators of other space types are within a preset range, the sky view factor and the green view factor of the open public space are higher than those of other space types and the excess amount reaches a first preset threshold, and the building view factor of the dense commercial block is higher than that of other space types and the excess amount reaches a second preset threshold.

[0028] Furthermore, it also includes calculating the theoretical time cost evaluation of visiting the public space according to the quality evaluation index data set of each public space in the target area, the public space type classification result and the functional facility density in each public space, and using the theoretical time cost of visiting the public space to evaluate the potential service capacity of the public space. The calculation expression of the theoretical time cost of visiting the public space is:

[0029]

[0030] in, represents the theoretical time cost of visiting the i-th public space, , They represent the minimum commuting time and maximum commuting time to visit public spaces, represents the potential service attractiveness of the public space of the i-th public space, represents the basic weight calculated according to the quality and function of public space, The ontological feature values ​​of each public space calculated using the quality assessment indicator dataset and functional facility density, Represents the weight of the nth ontology feature value.

[0031] Furthermore, it also includes determining optimization strategies based on the evaluation results of the potential service capacity and actual attractiveness of each public space, including:

[0032] Evaluate the actual attractiveness of the public space according to the actual flow of people in the public space to obtain an actual attractiveness evaluation result, wherein the actual flow of people data includes the number of visitors and residents;

[0033] For the first area where the actual attraction of the public space is higher than the potential service capacity, the POI mix and the pedestrian level are optimized by adjusting the types of POIs and the number of pedestrian facilities. For the second area where the actual attraction of the public space is lower than the potential service capacity, the public services and visual effects are optimized according to the corresponding characteristic indicator values ​​in the quality evaluation indicator data set.

[0034] A computer device comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0035] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed.

[0036] Compared with the prior art, the advantages of the present invention are as follows: the present invention establishes a multi-dimensional evaluation system from the quality and functional dimensions of public spaces, and analyzes the quality dimensions of public spaces from the three dimensions of global vision, global space and local vision at the same time, and combines the semantic segmentation model to realize the automatic extraction of public space features, which can comprehensively and accurately reveal the external morphological characteristics of the space, so that the comprehensive quality and functional dimensions can accurately reflect the potential use effects of public spaces, forming a complete matrix from image perception to material space quality and functional evaluation, and at the same time, the features extracted from the three dimensions of global vision, global space and local vision are further used to perform spatial clustering analysis through a deep learning model, which can also realize category detection of public spaces and effectively quantify the categories of public spaces, so as to facilitate the corresponding optimization and transformation of public spaces of different categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a detailed flowchart of a method for evaluating the quality of a public space based on multi-dimensional features according to an embodiment of the present invention.

[0038] Figure 2It is a distribution effect diagram of Baidu Street View data sampling points and on-site shooting sampling areas in a specific application embodiment of the present invention.

[0039] Figure 3 It is a schematic diagram of the effect of realizing the projection correction process in a specific application embodiment of the present invention.

[0040] Figure 4 It is a schematic diagram of the effect of merging fragmented segmentation results in a specific application embodiment of the present invention.

[0041] Figure 5 It is a schematic diagram of the global visual feature index result obtained in a specific application embodiment of the present invention.

[0042] Figure 6 It is a depth estimation sample diagram used in the calculation of the global spatial feature index obtained in a specific application embodiment of the present invention.

[0043] Fig. 7A It is a sky view factor result diagram in a visualization integration example obtained in a specific application embodiment of the present invention.

[0044] Figure 7B It is a green view factor result diagram in a visualization integration example obtained in a specific application embodiment of the present invention.

[0045] Figure 7C It is a result diagram of building view factors in a visualization integration example obtained in a specific application embodiment of the present invention.

[0046] Fig.7D It is a visual entropy result diagram in a visualization integration example obtained in a specific application embodiment of the present invention.

[0047] Figure 8 It is a schematic diagram of the detection results of the facility elements of the public space obtained in a specific application embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0049] It is difficult to accurately and comprehensively characterize the quality status of public spaces by considering a single spatial visual feature. In addition, the quality characteristics of different categories of public spaces are different. Simple quality ratings cannot effectively quantify the categories of public spaces. Starting from the quality and functional dimensions, the present invention adopts a "data-driven, model-supported, multi-dimensional evaluation" approach to establish a multi-dimensional evaluation system from the quality and functional dimensions of public spaces. The functional dimension is used to describe the quality status of public spaces from the perspective of use, so as to analyze the efficiency, convenience, and functional adaptability of public spaces. The quality dimension is used to focus on the physical attributes of public spaces, such as visual features, landscape composition, and layout complexity. At the same time, for the quality dimension of public spaces, a global perspective is used to evaluate the quality of public spaces. It analyzes the three dimensions of global vision, global space and local vision, and combines the semantic segmentation model to realize the automatic extraction of public space features, which can comprehensively and accurately reveal the external morphological characteristics of the space, so that the comprehensive quality and function dimensions can accurately reflect the potential use effect of the public space, forming a complete matrix from image perception to material space quality and function evaluation. At the same time, the features extracted from the three dimensions of global vision, global space and local vision are further used for spatial clustering analysis through deep learning models, which can realize the classification of public spaces based on the characteristics of the data itself, realize the category detection of public spaces, and effectively quantify the categories of public spaces, so as to achieve corresponding optimization and transformation for public spaces of different categories.

[0050] The following combination Figure 1 The public space quality assessment method based on multi-dimensional features of the present invention is described in detail, and the steps include:

[0051] Step S01: Acquire panoramic images of different types of public spaces in a target area to form a panoramic image dataset.

[0052] In this embodiment, the panoramic image data set includes panoramic images corresponding to the target public space obtained from map data and panoramic images collected on-site at multiple collection points in the target public space, that is, multi-source panoramic image data of the target research area is obtained by obtaining map data and taking on-site panoramic photos.

[0053] Specifically, we can obtain a panoramic photo dataset of four types of public spaces, namely streets, squares, green spaces and parks, in the study area by crawling Baidu Street View and taking on-site panoramic photos. The dataset includes information such as the crawled or taken photo image, photo time, photo number, photo longitude and latitude, and the category and scale of the public space to which it belongs. The above data information is then integrated through a geographic information system (GIS) to output a comprehensive database in the form of vector data.

[0054] As an optional implementation, the detailed steps of obtaining a panoramic image dataset by crawling Baidu Street View and taking on-site panoramic photos are as follows:

[0055] Step S101. Obtain multi-source geo-tagged photo data in the target study area: obtain the longitude and latitude of the public space through the Baidu Map API geocoding function, combine the Open Street Map (OSMs) vector data and the GIS pre-stored street view coordinates, use the coordinate conversion service to convert the WGS84 (World Geodetic System-1984 Coordinate System) coordinate system to the BD09MC coordinate system and remove abnormal data, and finally obtain the coordinate information of the street view collection points in the target area in batches by reading the CSV file.

[0056] Step S102. Based on the coordinate information, call Baidu API to obtain the street view image ID, obtain the panoramic image of the target public space in the Baidu map street view service through the web crawler technology, and control the request interval through the timer to prevent frequency blocking. After using the thread pool to batch download the image block byte data, use the image stitching code to merge and generate an equirectangular projection panorama.

[0057] Specifically, based on the coordinate information of the point, the street view image ID of the corresponding location can be obtained through the Baidu Map API, and then the uniform resource locator (URL) parameters sent by the standard HTTP request can be set, the download mode can be set to panoramic, and the interval of the request can be controlled by timing during the download process to avoid server blocking due to too fast request frequency. The thread pool is used to batch download blocks of different areas of the street view image and return their byte data lists. Based on the list, multiple image blocks are merged into a complete street view image through the image stitching code, and the panoramic image of the equirectangular projection is output.

[0058] Step S103. Check data integrity and create a list of areas to be supplemented, which are supplemented through Baidu web page manual crawling tools and on-site collection of multiple typical spaces.

[0059] Specifically, when checking the integrity of the batch-acquired Baidu Street View image data, a list of areas to be supplemented is established, and supplementary data is obtained in two ways: one is to use development tools to manually crawl image data on the Baidu Map web page; the other is to select multiple typical public spaces that lack street view data as supplementary collection locations, use panoramic cameras for on-site collection, and output images as equirectangular projection panoramas. The rules for the layout of collection points are as follows: for linear spaces, a collection point is set every 50 meters (configurable) along the axis; for square-type open spaces, collection points are set at the center of the space and at the surrounding boundaries. The distance between the boundary collection point and the boundary structure should ensure that the boundary elements are clearly identifiable; for key nodes (including but not limited to landmark buildings, road intersections, etc.), additional collection points are set for key collection.

[0060] Step S104. Metadata (coordinates / image ID, etc.) of the street view image data are recorded, and metadata information such as coordinates and image ID of each image is recorded in a data table. Data is filtered according to the boundary of the study area, and data anomalies (image acquisition failure, coordinate conversion error, etc.) are reviewed and corrected. Then, consistency checks are performed on the groups according to the collection source to ensure the comparability of data between groups. Finally, the images are processed through the Python OpenCV library, and Baidu Street View is selectively cropped to remove non-target areas (such as collection vehicles). The images collected on site are resampled to a unified pixel specification to complete the standardization of multi-source data.

[0061] Taking the above steps to collect panoramic images as an example, the distribution of Baidu Street View data sampling points and on-site shooting sampling areas is shown below: Figure 2 shown.

[0062] It is understandable that other methods may be used to acquire a panoramic image of the target research area according to actual application requirements, and other preprocessing / standardization methods may be used for the panoramic image data.

[0063] Step S02: Perform semantic segmentation on each panoramic image in the acquired panoramic image data set using a semantic segmentation model to obtain a semantic segmentation result, and process the semantic segmentation result to obtain a semantic segmentation processing result, wherein the semantic segmentation processing result includes the area ratio of each type of semantic segmentation element.

[0064] By performing semantic segmentation on an image, the semantic meaning of different areas in the image can be obtained. After processing, the information such as the area of ​​the semantic segmentation elements and the corresponding mapping relationship between the segmented areas and the feature categories can be obtained. This information can be used to calculate various feature indicators to evaluate the quality of public space. The semantic segmentation element area is the area of ​​all feature categories in the image. The feature categories include sky, building, green space, traffic elements, public elements, etc. The corresponding mapping relationship between the segmented areas and the feature categories is the corresponding relationship between each segmented area and each feature category. After determining the segmented image, the area proportion of each type of semantic segmentation element can be obtained through the pre-constructed mapping dictionary between color and feature categories.

[0065] As an optional implementation, the following steps may be used to perform semantic segmentation on the panoramic image and process the semantic segmentation result:

[0066] Step S201: Obtain a semantic segmentation result by inputting the acquired panoramic image dataset into a semantic segmentation model.

[0067] As an optional implementation, a PSPNet (Pyramid Scene Parsing Network) semantic segmentation model can be used for semantic segmentation. The model uses ResNet-101 as the backbone network, with a downsampling rate of 8. The deep learning model with 101 layers (configurable) of convolutional layer structure obtained through iterative training performs large-scale, objective semantic segmentation processing on the panoramic image content obtained in step S01, outputs a segmentation mask, and then generates a color area overlay image based on the mask as the segmentation result. The PSPNet model extracts contextual information at different scales by introducing a pyramid pooling module (PPM), which can better understand the global and detailed features of the image, solve the multi-scale problem in scene parsing, and stably and reliably implement the segmentation task of global information modeling and the parsing of complex scenes.

[0068] As an optional implementation, the semantic segmentation model is trained using 4 GPUs with a batch size of 4 based on the ADE20K and Cityscapes datasets. The ADE20K dataset contains more than 20,000 images, covering 150 different categories of objects and scenes. All images have been semantically labeled in detail to ensure the model's good performance in multiple scenes and object classifications. Cityscapes provides pixel-level annotations of 50 urban road scenes, which can be used for image datasets of semantic segmentation of urban scenes. This dataset annotates the semantic labels of each pixel in each image, and the annotation information covers different types of objects (such as cars, buildings, trees, etc.) and scene parts (such as the sky, the ground, etc.).

[0069] As an optional implementation, a pixel-level semantic segmentation mask is generated for the panoramic image by loading a pre-trained model, and a specific RGB value is assigned to each semantic category to form a color segmentation area. Specifically, the pre-trained model is applied to the input panoramic image to generate a pixel-level semantic segmentation mask file, and a specific RGB color value is assigned to each semantic category to generate a color segmentation area. Finally, the color segmentation area is layered with the original input image, and the superposition result is saved in PNG format. Compared with the traditional method of directly reading the mask file, the use of PNG format to output the color segmentation area image can ensure that the color information in the image is completely preserved under lossless compression in the subsequent projection correction process, thereby ensuring the data accuracy during color extraction and analysis in the subsequent steps.

[0070] Step S202: Taking the semantic segmentation result as the segmentation image, the Hammer-Aitoff equal product projection transformation method is used to perform projection correction on the segmentation image, and the area of ​​each type of semantic segmentation element in the segmentation image is calculated according to the result of the projection correction.

[0071] This embodiment uses the Hammer-Aitoff equal-area projection transformation method to correct the equirectangular projection segmentation image into an equal-area plane image, so that the area of ​​various semantic segmentation elements can be accurately obtained. The corrected image can then generate a corresponding mapping result between the segmentation area and the element category by indexing the metadata of the data set (such as the above-mentioned ADE20K data set). In traditional street view analysis methods, equirectangular projection is usually used directly for processing, but equirectangular projection will produce serious geometric deformation in the area around the image, especially in high-latitude areas, which will cause objects at the edge of the image to be stretched or compressed, and then cause deviations in the spatial distribution and quantitative evaluation of the analysis results. This embodiment uses the Hammer-Aitoff equal-area projection correction method to correct the equirectangular projection segmentation image, which can effectively reduce image deformation and maintain the consistency of area ratio, thereby more realistically reflecting the spatial distribution characteristics of the street scene, solving the problem of edge deformation of panoramic images, and improving the accuracy of spatial analysis.

[0072] Furthermore, considering that the amount of data of the street view image itself is large, the interpolation calculation of the projection correction may be slow. Step S202 performs projection correction on the segmented image by using the Hammer-Aitoff equal product projection transformation method. Specifically, the following steps can be used:

[0073] Step S221. Perform image preprocessing and coordinate mapping preparation: establish a grid system of standardized longitude and latitude coordinates of the panoramic image.

[0074] Specifically, the longitude and latitude coordinate grid is a spherical coordinate system, and the longitude range can be set within the interval [-π,π], and the latitude range can be set within In the interval, the Hammer-Aitoff projection plane coordinate system is constructed based on the projection sphere radius R, and the plane coordinate range after projection is determined as the x-axis , the y-axis range is In actual processing, the specified ratio (such as 10%) can be extended to the x-axis , y-axis To accommodate the interpolation margin, to avoid invalid data filling at the edge of the projected image, and to generate a dense longitude and latitude grid matrix through gridding processing;

[0075] Step S222. Execute coordinate system conversion: perform Hammer coordinate system conversion on the standardized longitude and latitude coordinates of the panoramic image to convert them into x and y coordinates of the Hammer-Aitoff projection plane coordinate system.

[0076] Specifically, the longitude offset of the standardized longitude and latitude coordinates of the panoramic image established in step S221 relative to the central meridian is calculated, and the spherical coordinates of the equidistant cylindrical projection are converted to the plane coordinates of equal area based on the Hammer-Aitoff projection formula. The formula for converting the standardized longitude and latitude coordinates of the panoramic image into the x and y coordinates of the Hammer-Aitoff plane coordinate system is as follows:

[0077]

[0078] Among them, z is an intermediate calculation quantity used to normalize the values ​​of x and y to achieve equal area projection. latitude represents the standardized latitude of the panoramic image, and longitude represents the standardized longitude of the panoramic image.

[0079] Furthermore, the computational efficiency can be optimized by using just-in-time compilation technology, and invalid coordinate points can be screened out by threshold setting and conditional judgment to further ensure the projection accuracy.

[0080] Step S223. Based on the coordinate system conversion result of step S222, the segmented image is reprojected.

[0081] To improve processing accuracy and speed, this embodiment first separates the RGB channels of the input panoramic segmentation image, independently performs Hammer coordinate system transformation on each color channel in the manner of step S222, uses a bilinear interpolation algorithm to grid-resample discrete data points to fill in missing data, and merges the three channels after processing to generate a complete projected segmentation image, thereby realizing reprojection processing of the segmented image.

[0082] Furthermore, batch conversion can be accelerated in parallel by dynamically optimizing the thread pool to further improve projection efficiency. For example, by enabling CUDA kernel parallel computing for each color channel, using a bilinear interpolation algorithm to achieve sub-pixel resampling, and optimizing the GPU memory histogram, the three-channel processing latency can be reduced by about 30%.

[0083] Step S224. Assign a corresponding category index to each semantic segmentation element category in advance, and construct a bidirectional mapping relationship between RGB color and category index, read the projected segmented image, and convert the projected segmented image in the RGB color space into the corresponding category index pixel by pixel according to the pre-constructed bidirectional mapping relationship between RGB color and category index to obtain a category index map, which can be used to calculate the area of ​​each segmentation element later.

[0084] Traditional street view analysis methods usually adopt the method of directly reading the mask file in the semantic segmentation result to obtain the area proportion of each type of segmentation element. This embodiment can accurately calculate the area of ​​the segmented area by first performing semantic segmentation, then projecting, and finally reading through color mapping. This embodiment also includes creating a zero matrix of the same size as the original image as a category annotation container, replacing the pixel value of the corresponding position with the corresponding category index through pixel-by-pixel color matching, and realizing the conversion from visual representation to semantic annotation.

[0085] Specifically, a standardized semantic segmentation color mapping system can be built through Python. A bidirectional mapping system of RGB color-category index for 150 types of scene elements can be built based on the ADE20K dataset. The scene element dictionary containing 150 categories in the ADE20K dataset is indexed. Each category is identified by a unique RGB color value. A corresponding category index is assigned to each semantic segmentation feature category to form a bidirectional mapping relationship from color value to category index. The PIL library is used to batch read PNG segmented images and convert them into NumPy arrays. Pixel-level semantic classification conversion is performed. The segmented image in RGB color space is converted into a category index map using a predefined color mapping relationship. At the same time, a zero matrix of the same size as the original image is created as a category annotation container. The pixel values ​​at the corresponding positions are replaced with the corresponding category indexes through pixel-by-pixel color matching to achieve accurate conversion from visual representation to semantic annotation. Finally, the mapped category index map can be used to accurately calculate the area proportion of various types of semantic segmentation elements.

[0086] Take the above steps to implement projection correction in actual public space as an example. Figure 3 As shown, Figure 3 (A) corresponds to Baidu Street View images, (B) corresponds to on-site shooting and Baidu interior view images, (C) corresponds to cropping the area of ​​the street view collection vehicle, (D) corresponds to the semantic segmentation results of Baidu Street View images, (E) corresponds to the semantic segmentation results of on-site shooting and Baidu interior view images, (F) corresponds to the projection correction results of Baidu Street View images, and (G) corresponds to the projection correction results of on-site shooting and Baidu interior view images.

[0087] Step S03. Calculate the global visual feature index, local visual feature index and global spatial feature index of each public space according to the result of semantic segmentation processing, and generate a quality assessment index data set for each public space, so as to evaluate each public space from the global vision, local vision and global spatial dimensions. The global visual feature index is the overall visual perception feature that characterizes the public space, the local visual feature index is the feature that characterizes the specific entity in the perception process of landscape elements, and the global spatial feature index is the feature that characterizes the scale and spatial form of the public space.

[0088] This embodiment analyzes the quality indicators of public spaces by analyzing multi-source panoramic image data, extracting the visual and spatial features of public spaces, and processing different types of visual data through standardized methods, so as to accurately identify indicators such as global vision, global space, local space, and the number of categories of functional facilities in the space, thereby constructing an efficient spatial interaction intelligent measurement system; then, the public space evaluation results obtained based on the quality indicators and functional facilities are spatially integrated and superimposed, and the spatial demand, time cost, and capacity utilization of the public space are further considered to establish an attractiveness model of the public space to analyze the potential of public spaces to provide services, so as to facilitate the formation of quality improvement and optimization plans for different types of public spaces. Finally, the potential service capacity and actual attractiveness evaluation results of each public space are used to form a classified evaluation result to determine the corresponding optimization strategy and differentiated transformation plan.

[0089] First, this embodiment divides the contents of different stages of environmental perception and cognition into three dimensions, and establishes a public space quality assessment index system to unify multi-dimensional assessment variables and meet the dynamic and granular requirements of public space. The index system is divided into three dimensions according to the level of perception and cognitive processing: (1) global visual feature index; (2) global spatial feature index; (3) local visual feature index, forming a comprehensive assessment framework covering spatial perception, landscape elements and spatial form.

[0090] Specifically, the global visual feature indicators include sky view factor, building view factor, green view factor, visual entropy, skyline fractal dimension and color properties, among which the sky view factor uses the proportion of sky pixels to measure the density and sense of closure of urban buildings, the building view factor uses the proportion of building pixels to measure the sense of closure and building density of urban space, the green view factor uses the proportion of green vegetation pixels to measure the degree of greening, the visual entropy is used to measure the complexity of information in the image, the skyline fractal dimension is used to characterize the spatial complexity of the building outline, and the color properties include color brightness, contrast, saturation, etc.

[0091] Specifically, local visual feature indicators include the proportion of natural element classes, the proportion of traffic elements, the proportion of public facilities, the proportion of safety facilities, the proportion of human and vehicle activities, the proportion of street walls, the proportion of building signs, the proportion of street-facing windows and the proportion of designated elements. The proportion of natural element classes is used to characterize the proportion of natural elements in the image, the proportion of human and vehicle activities is used to characterize the proportion of the number of people and vehicles in the image, and the proportion of designated elements is the proportion of elements in the segmented image results that is greater than a preset proportion threshold or the proportion of elements of a specified type, such as an element proportion greater than 5% or other element proportions selected by mainstream research.

[0092] Specifically, the global spatial characteristic indicators include the proportion of vertical elements, walking width, block length ratio, H / W ratio of street canyons, spatial connectivity and depth map spatial ratio. The proportion of vertical elements is used to characterize the relationship between building height and block scale, walking width is the width of the street, block length ratio is the ratio of block length and width, the H / W ratio of street canyons is the ratio of building height to street width, spatial connectivity is used to characterize the traffic and connectivity between spaces, and the depth map spatial ratio is the ratio of different space types in the depth map.

[0093] The specific configurations of the above-mentioned global visual feature indicators, local visual feature indicators and global spatial feature indicators are shown in Table 1.

[0094] Table 1: Public space quality assessment index table

[0095]

[0096] As an optional implementation, in order to calculate the above-mentioned feature indices, the semantic segmentation processing results obtained in step S02 are respectively input into different calculation modules, and the global visual feature index value (such as sky view factor, building view factor, etc.) is calculated by element proportion, and the local visual feature index is calculated by classification or clustering.

[0097] The standard values ​​(such as natural elements, traffic elements, proportion of street facilities, etc.) are used to calculate the global spatial characteristic index values ​​(such as proportion of vertical elements, walking width, block length, etc.) through the Dinov model for depth estimation.

[0098] Taking visual entropy calculation as an example, we can read panoramic photos, use OpenCV and sklearn to capture the lighting and color features (brightness / contrast / saturation, etc.) of the panoramic image, convert the image into a grayscale image through the Pillow library and divide it into n areas based on the grayscale, and then calculate the visual entropy based on the grayscale distribution probability. The formula is as follows:

[0099] Entropy (4)

[0100] in Represents the gray value of the probability of each region appearing The higher the entropy value, the more complex the image.

[0101] Taking the fractal dimension of the skyline as an example, we can read the color area image obtained after the panoramic image is segmented, calculate the fractal dimension of the skyline to measure the complexity of the city skyline, and use it to characterize the spatial complexity of the building outline. The higher the fractal dimension, the more complex and irregular the skyline is. A good fractal structure helps humans perceive space naturally, quickly, and pleasantly. The sky contour line can be extracted from the semantic segmentation result. The box counting method can be used to calculate the fractal dimension. The contour is covered by grids of different sizes, and the rate of change of the number of covered grids as the grid is reduced is counted. The logarithmic slope is fitted as the fractal dimension to quantify the spatial complexity of the skyline. The box counting method essentially measures the growth rate of pattern complexity on a reduced scale. First, the pattern O is covered with a square grid box, and then the side length of the box is continuously reduced and the process is repeated. The number of grids N (ε) and the inverse of the side length of the box 1 / ε are recorded in each repeated step. The slope of the fitted line on the logarithmic graph is the fractal dimension D (O). The calculation formula can be expressed as follows:

[0102] D(O)= (5)

[0103] As an optional implementation, pixel statistics and data organization can be performed on the color area image after equal area correction output in step S02, and statistical analysis can be performed on the converted category index map to calculate the number of pixels appearing in each semantic category in the image, and record the image file name and total number of pixels and other meta-information, and organize the statistical results into structured data. For example, the number of pixels in each category in the semantic segmentation result is counted, the file name and total number of pixels are integrated, and the name and location of the public space, as well as the photo number, photo location, shooting time and other meta-information are integrated into the pandas data frame and output as a CSV file.

[0104] For complex scenes in partial panoramic photos, the model may segment a single object into multiple parts (over-segmentation) or merge multiple objects into one area (under-segmentation). This embodiment optimizes the results by post-processing the 19 types of elements output by cityscape and the 150 types of element classifications output by the ADE20K dataset through labeling and other means, and merges the fragmented elements identified as paths, sidewalks, floors, runways, roads, soil, etc. in the semantic segmentation results into sidewalks, roads and all road surfaces. For example, paths / sidewalks are merged into road surface classes, and grass / trees are merged into vegetation classes, thereby solving the over-segmentation / under-segmentation problem and optimizing the consistency of classification results. In a specific application embodiment, the merged result obtained by using the above method on the fragmented segmentation results is as follows: Figure 4 As shown, Figure 4 The upper middle part corresponds to the original result of over-segmentation, and the lower middle part is the example result after merging and annotation.

[0105] This embodiment further calculates the proportion of each element according to the number of pixels appearing in each semantic category and the total number of pixels in the image, and then calculates the values ​​of the sky, building and green view factors. The distribution differences of the three types of factors in the public space can also be visualized through a bar chart. The specific formulas of the sky view factor, building view factor and green view factor are as follows:

[0106] Sky View Factor 100% (6)

[0107] Building View Factor 100% (7)

[0108] Green View Factor 100% (8)

[0109] As an optional implementation, the unlabeled self-distillation v2 model (DINOv2) is used to estimate the depth of the panoramic image and calculate the global spatial feature index results. The depth map pixel value generated by the model ranges from 0 to 255, corresponding to the relative distance of the landscape element from the observation point. Based on the generated depth map, the average depth and depth standard deviation of the landscape scene are calculated to evaluate the global spatial characteristics of the landscape. The DINOv2 model is a self-supervised learning model based on the Transformer framework. It can use pre-trained models for reasoning under zero-sample conditions. Compared with traditional on-site measurement methods and deep learning models that require a large amount of training data, it can not only achieve automated and efficient evaluation of public space perception, but also handle the complex geometry of public spaces.

[0110] For example, the DINOv2 model is used to generate a panoramic depth map (pixel values ​​0-255 represent relative distances), the average depth and standard deviation are calculated, and the four regions are divided into foreground FG (≤64), mid-ground MG (65-128), background BG (129-192), and distant background EBG (≥192) according to the depth value. The proportion of each region is calculated as follows:

[0111]

[0112] in, Indicates pixels with depth values ​​≤ 64, Indicates pixels with depth values ​​in the range of 65-128. Indicates pixels with depth values ​​in the range of 129-192. Indicates pixels with depth values ​​≥ 192, and N indicates the total number of pixels.

[0113] Furthermore, the mean, median and quartile of each indicator (visual entropy, fractal dimension, view factor, depth partition) can be calculated, and the statistical distribution can be displayed through box plots. The indicator results can be mapped to spatial locations in combination with the GIS graded symbol diagram method, that is, the graded symbol diagram method is used in GIS to locate the visual entropy, skyline fractal dimension and three types of view factors of each public space in space, so as to visualize their spatial differences and reveal the spatial differences of the visual characteristics of public spaces. Taking the above steps to calculate the characteristic indicators of Zhuzhou public space as an example, the global visual characteristic indicator results are as follows: Figure 5 As shown, Figure 5 (A) corresponds to the bar graph of sky, green view, and building view factors, (B) corresponds to the box graph of visual entropy, and (C) corresponds to the box graph of skyline fractal dimension. The depth estimation sample used in the calculation of global spatial feature indicators is shown in the figure. Figure 6 As shown in the figure, the final quality assessment result is visualized as follows: Figure 7A to Figure 7D As shown, Fig. 7A The corresponding sky view factor result is Figure 7B The corresponding result is the green view factor. Figure 7C The corresponding building view factor result is Fig.7D The corresponding result is the visual entropy.

[0114] This embodiment combines the identity of the top-down theoretical system with the objective needs of dynamic and granular public space evaluation. By constructing a comprehensive evaluation framework of three dimensions: global vision, global space, and local vision, a complete multi-dimensional evaluation system is established. It can unify samples and variables of different dimensions, establish a comprehensive matrix from objective material space to subjective experience evaluation, and form a quantitative evaluation system from spatial material form to specific spatial optimization of urban residents' lives, thereby achieving an all-round measurement of public space quality. The evaluation system has good adaptability and can be used for the evaluation of different types of public spaces.

[0115] Step S04: Detect the categories of public space functional facilities based on the semantic segmentation processing results, and evaluate the functions of the public space based on the detection results.

[0116] On the basis of semantic segmentation, this embodiment further detects the categories of public space functional facilities to achieve automatic detection and positioning of public space facilities. The density of functional facilities can be obtained based on the detection results of functional facility categories, and the public space can be quantitatively evaluated from the functional dimension, thereby providing a quantitative basis for functional layout optimization and effectively solving the problem of insufficient adaptability of traditional evaluation methods in complex scenarios.

[0117] As an optional implementation, the Grounding DINO model can be used to detect public space facilities to identify specific functions of public spaces, such as entertainment, sports, public services, etc. A set of facility text prompts is pre-established, including but not limited to text descriptions of public facilities such as seats, trash cans, runways, traffic signs, zebra crossings, and bicycle racks. The public space panoramic image is input into the Grounding DINO model for target detection and positioning. The facility bounding box is entropy drawn on the original panoramic image and the category label is annotated. The detection result with annotation information is output, such as an annotation map. The annotation information includes facility category, bounding box coordinates, detection confidence, and location information. The Grounding DINO model combines DINO (Transformer-based target detection model) and Grounding mechanism to achieve target detection by fusing image and text features, and can detect corresponding objects in the image based on text descriptions.

[0118] Preferably, the parameter combination of setting box_threshold (bounding box screening threshold) to 0.4 and text_threshold (text matching threshold) to 0.25 in the Grounding DINO model can obtain the best performance. Taking the above steps to detect the facility elements of a public space as an example, the detection results are as follows: Figure 8 shown.

[0119] This embodiment makes full use of deep learning models such as PSPNet, DINO v2, and Grounding DINO, and realizes automatic extraction of public space features through computer vision technology, which can significantly improve the evaluation efficiency, and the evaluation results are repeatable and objective, avoiding subjective bias in manual evaluation.

[0120] Step S05: Clustering the quality assessment index data sets of each public space to implement spatial clustering analysis and obtain the public space type of each public space.

[0121] Taking into account the different representations of characteristic indicators of different types of public spaces, this embodiment clusters the quality assessment indicator data sets of each public space, mines the data ontology feature structure through dynamic data clustering, and automatically classifies the categories of urban public spaces. Compared with the traditional simple high or low or numerical grading, it can perform differentiated evaluation according to the quality and functional characteristic values ​​of different public spaces, thereby realizing pixel-level feature extraction and quantitative analysis, which can provide a quantitative basis for space optimization and facilitate the provision of one-location-one-policy optimization solutions for different types of public spaces.

[0122] As an optional implementation, the K-means clustering optimization algorithm based on adaptive threshold is used to cluster the quality assessment index data set. The types of public spaces include mixed public spaces, open public spaces and dense commercial blocks. Each type of space has a unique combination of index characteristics, which realizes the accurate automatic classification of public spaces. The difference between the characteristic index values ​​of the mixed public space and the characteristic index values ​​of other space types is within the preset range, that is, the characteristic index values ​​are in a balanced state. The sky view factor and green view factor of the open public space are higher than those of other clusters and the excess reaches the first preset threshold. The building view factor of the dense commercial block is higher than that of other space types and the excess reaches the second preset threshold. The first preset threshold and the second preset threshold can be configured according to actual needs. The core idea of ​​the K-means clustering optimization algorithm based on adaptive threshold is to divide the sample points into several clusters through iterative optimization, so that the similarity of the sample points within the cluster is maximized and the difference between clusters is maximized. During the evaluation, the threshold and scale of the cluster are dynamically adjusted according to the distribution of the data through the adaptive quantile method, so as to effectively handle the dynamically updated data of the public space evaluation. Through the dynamic data-driven K-means clustering spatial classification method, accurate and automatic classification of urban public spaces can be completed.

[0123] Specifically, the sklearn library of Python can be used to select multiple public space data for sample experiments, determine the optimal K value to be 3, select 3 sample points as the starting cluster center in the multidimensional index space, and calculate the distance from each sample point to the cluster center. The Euclidean distance calculation method is used to determine the distance between each public space sample point and each cluster center. According to the distance calculation result, each public space is divided into the cluster category with the closest distance, and then the mean of all sample points in each cluster is recalculated. The mean point is used as the new cluster center to divide the nearest distance cluster and calculate the mean of the samples in the cluster until the cluster division tends to be stable, that is, the position change of the cluster center does not exceed the preset threshold. Finally, each type of space has a unique combination of high-weight features to form differentiated spatial quality features. High-weight features refer to the fact that in the clustering process, a certain feature value of all samples in a certain cluster is significantly higher than that of other clusters (determined by mean comparison or variance analysis). These features have a high degree of distinction within the cluster and can represent the core environmental attributes of the cluster. The final results can be divided into three categories: mixed public space, open public space and intensive commercial blocks.

[0124] For example, the ecological indicators of open public spaces are more prominent, such as the characteristic mean of the sky view factor is 0.17~0.36, and the characteristic mean of the green view factor is 0.29~0.31, which is significantly higher than other clusters; the characteristic indicators of mixed public spaces are balanced, such as the characteristic mean of the sky, building and green view factors and the proportion of various elements compared with other clusters, and there are no obvious prominent indicators. Dense commercial blocks are characterized by high building density and transportation facilities indicators. The characteristic mean of the building view factor of this cluster is 0.24~0.37, which is significantly higher than that of open public spaces (characteristic mean is 0.05~0.18) and transportation facilities (0.07~0.09), while the characteristic mean of open public spaces is close to 0.

[0125] Step S06. Evaluate the public space attractiveness of each public space in the target area according to the quality assessment index dataset of each public space, the public space type classification result, and the functional facility density in each public space.

[0126] This embodiment combines the quality and function of each public space, calculates the service level and theoretical visit time cost of the public space according to the public space quality and function evaluation results, so as to analyze its potential service capacity on a site-by-site basis, thereby facilitating the determination of specific space optimization plans and solving the dilemma of the sameness of all sites in traditional planning and management.

[0127] As an optional implementation, the calculation results of steps S01 to S05 can be integrated and input into a geographic information system (GIS), and spatial distribution feature maps of different indicators can be generated through layered and colored vector symbols. The evaluation results can be intuitively displayed on the GIS platform, revealing the quality characteristics of public spaces in a data-driven manner, covering the planning, construction and governance cycle of public spaces, thereby providing accurate improvement suggestions for the optimization of urban public spaces.

[0128] As an optional implementation method, the public space evaluation results obtained based on quality indicators and functional facilities are spatially integrated and superimposed for analysis, and the public space attraction model is established by considering the spatial demand, time cost and capacity utilization of the public space to analyze the potential of public space to provide services, so as to form quality improvement and optimization plans for different types of public spaces. Specifically, the theoretical time cost evaluation of visiting public space is calculated based on the spatial demand, time cost and capacity utilization of public space, and the theoretical time cost of visiting public space is used to evaluate the potential service capacity of public space. The calculation process is as follows:

[0129] First, the basic weights of public space quality and function are calculated , the calculation formula is as follows:

[0130] (13)

[0131] in: represents the nth ontological eigenvalue, is the basic weight of the ith public space, is the weight coefficient corresponding to the nth ontological eigenvalue, which is determined by the capacity of the public space. The ontological eigenvalue is a variety of indicator characteristics used to describe the service provision capacity of urban public spaces. It contains the quality characteristic results of public spaces and the detection results of functional facilities. This information jointly determines the potential of public spaces to provide specific public services. For example, for leisure services, global and local visual characteristic indicators such as sky view factor, building view factor, green view factor, visual entropy, and proportion of safety facilities in the quality of public spaces, as well as the category and density of functional facilities, will affect its ability to provide leisure services.

[0132] Specifically, the ontological feature values ​​include the quality evaluation indicators of the three dimensions of global vision, local vision and global space calculated in the above step S03 and the density of functional facilities detected in step S04, while taking into account the surrounding population density and population composition characteristic infrastructure, surrounding POI density (daily life, work, special functions), etc., and then calculating the attractiveness model of the public space. For example, = the result of the calculation of the sky view factor, =Proportion of transportation facilities, =Sports facility density.

[0133] Then, standardization is performed to normalize the weights of all public spaces to a uniform interval, so that the attractiveness of public spaces is calculated as a weight factor, and the formula is as follows:

[0134] (14)

[0135] in, represents the potential service capacity of the public space of the ith public space, The weight value of the public space with the smallest basic weight, Indicates the weight value of the common space with the largest base weight.

[0136] Then calculate the theoretical time cost of visiting public spaces, the formula is as follows:

[0137] (15)

[0138] in, represents the theoretical time cost of visiting the i-th public space, is the minimum commuting time to access public spaces, is the maximum commuting time, which can be obtained by analyzing the mobile phone signaling data of residents’ travel.

[0139] As shown in the above formulas (13) to (15), space demand, time utilization and actual capacity weights are considered. Space demand refers to the different spatial requirements of residents for different types of public spaces. Capacity utilization is the weight of the daily maximum capacity obtained by combining the capacity of each space with the time turnover. The space capacity is calculated by the average residence time of the public space to calculate the batches that can be accommodated every day, so as to calculate the weight of the daily maximum capacity by multiplying the capacity of each space by the visit batch. Finally, based on the above weights and indicators, the theoretical time cost of visiting the public space is calculated, and a model of the potential service capacity of the public space is established, so as to quantitatively evaluate the potential service capacity of the public space for urban residents. Furthermore, the actual attractiveness of the public space can be evaluated based on the actual flow data such as the number of visitors and residents in the public space to obtain the actual attraction evaluation result. The actual flow data can be obtained through data such as mobile phone signaling.

[0140] Step S07: Determine an optimization strategy based on the evaluation results of the potential service capacity and actual attractiveness of each public space.

[0141] This embodiment makes full use of the evaluation results of the potential service capacity and actual attractiveness of each public space, determines the corresponding optimization strategy, and realizes the accurate optimization of public space in the manner of "accurate diagnosis-scientific decision-making-dynamic regulation". Specifically, according to the classification evaluation results, a differentiated transformation plan is determined: for the first area where the actual attractiveness of the public space is higher than the potential service capacity, the POI mix and the walking level are optimized by adjusting the types of POIs and the number of walking facilities. For example, for areas such as pedestrian streets with a large number of actual visits by residents, the actual attractiveness is higher than the potential service capacity. The types of POIs around them can be appropriately increased to promote the growth of diversity, and pedestrian-friendly facilities can be added or rebuilt to provide walking and resting spaces for visitors with a large flow of traffic. For the second area where the actual attractiveness of the public space is lower than the potential service capacity, the public service and visual effects are optimized according to the values ​​of each characteristic indicator in the corresponding quality evaluation indicator data set. For example, according to the evaluation results of the global view and global space characteristic indicators, the walking width is increased, and vertical greening is performed to improve the evaluation value of the green view factor; for areas with a large area and a low visit density, the actual attractiveness is far lower than the potential service capacity, and the greening area can be appropriately reduced according to the evaluation results of the global visual characteristic indicators.

[0142] As an optional implementation method, the digital twin technology can be combined to realize the real-time dynamic determination of the optimization strategy. First, the dynamic monitoring capability of the deep learning model is used to establish an image spatiotemporal data perception system, and the pixel-level analysis of spatial elements is realized through the PSPNet network. Combined with the fine-grained feature recognition of Grounding DINO, a digital twin of spatial quality is constructed. Then, the three-dimensional spatial database corrected by Hammer-Aitoff projection is used to carry out multi-dimensional cross-analysis on the GIS platform. At the same time, an open source model-driven iterative optimization mechanism is established to couple the dynamically collected evaluation index data with the real-time data of people's visits and residence, generate a spatial performance optimization parameter package, and finally form a transformation plan covering visual experience improvement, visual space regulation, and functional patchwork, so as to achieve the precise improvement of public space quality.

[0143] Furthermore, the quality assessment results can be used to determine optimization strategies. For example, the visual ratio of sky buildings to green views, interface permeability, and visual richness can be optimized from a global visual dimension, spatial perception at human scale can be optimized from a spatial dimension, and specific sign information and public safety facilities can be optimized from a local visual dimension.

[0144] In summary, the present invention aims at the intelligent measurement of the pattern characteristics and spatial interaction of urban public spaces, and realizes the dynamic evaluation of the quality of public spaces based on the quality and functional dimensions. The quality of public space is analyzed through indicators of three dimensions: global space, global vision and local vision. The function of public space is analyzed through the detection of facility elements and the mining of public space ontology characteristics. A multi-dimensional quality evaluation system covering global vision, global space and local vision is established. In combination with deep learning models, large-scale and objective spatial feature extraction is achieved, forming an objective and comprehensive evaluation system based on multi-source image data and spatiotemporal big data. Combined with data mining and intelligent measurement technology, it can reveal information on the morphological characteristics, spatial distribution, public activity patterns, and human-land coordination of public space, forming a full-process, algorithmic, multi-data source, bottom-up public space quality evaluation system, which can accurately and comprehensively evaluate the quality of different types of public spaces, effectively mine the actual supply and demand of different types of public spaces, provide quantitative basis for the optimization and transformation of public spaces, accurately determine the layout of urban public spaces and the optimization plan of facility configuration, improve the quality and function of public spaces, and help to more effectively improve the functions and user experience of public spaces in urban space planning. At the same time, it can realize the automation and standardization of public space evaluation and significantly improve the evaluation efficiency.

[0145] This embodiment further provides a computer device, including a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method. It can be understood that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario by multiple devices cooperating with each other to complete. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing related programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other applications. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0146] This embodiment further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above method is implemented.

[0147] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for evaluating the quality of public space based on multidimensional features, characterized in that the steps include: Acquire panoramic images of different types of public spaces in a target area to form a panoramic image dataset; Performing semantic segmentation on each panoramic image in the acquired panoramic image data set using a semantic segmentation model to obtain a semantic segmentation result, and processing the semantic segmentation result to obtain a semantic segmentation processing result, wherein the semantic segmentation processing result includes an area ratio of each type of semantic segmentation element in the image; According to the area proportions of various semantic segmentation elements in the semantic segmentation processing results, the global visual feature index, the local visual feature index and the global spatial feature index of each public space are calculated respectively, so as to evaluate the quality of each public space from the global visual, local visual and global spatial dimensions respectively, and generate a quality evaluation index data set for each public space; The global visual feature index is a feature of the overall visual perception of the public space, the local visual feature index is a feature of the specific entity in the perception process of landscape elements, and the global spatial feature index is a feature of the scale and spatial form of the public space; Using a deep learning model to perform public space functional facility category detection on the semantic segmentation processing result, so as to evaluate the function of the public space according to the detection result; Clustering the quality assessment indicator data set to implement spatial cluster analysis and obtain the public space type of each public space; It also includes calculating the theoretical time cost evaluation of visiting the public space according to the quality evaluation index data set of each public space in the target area, the classification result of the public space type and the density of functional facilities in each public space, and using the theoretical time cost of visiting the public space to evaluate the potential service capacity of the public space. The calculation expression of the theoretical time cost of visiting the public space is: ； in, represents the theoretical time cost of visiting the i-th public space, , They represent the minimum commuting time and maximum commuting time to visit public spaces, represents the potential service capacity of the public space of the ith public space, represents the basic weight of the i-th public space, is the nth ontology eigenvalue calculated using the quality assessment indicator dataset and functional facility density, Represents the weight of the nth ontology feature value.

2. The public space quality assessment method based on multidimensional features according to claim 1 is characterized in that: The step of performing semantic segmentation on each panoramic image in the acquired panoramic image data set using a semantic segmentation model to obtain a semantic segmentation result, and processing the semantic segmentation result to obtain a semantic segmentation processing result includes: The semantic segmentation result is obtained by inputting the acquired panoramic image dataset into the semantic segmentation model; The semantic segmentation result is used as a segmentation image, and the segmentation image is projected and corrected by using a Hammer-Aitoff equal-product projection transformation method. The areas of various semantic segmentation elements in the segmentation image are calculated according to the projection correction result.

3. The public space quality assessment method based on multidimensional features according to claim 2 is characterized in that: The use of the Hammer-Aitoff equal-area projection transformation method to perform projection correction on the segmented image comprises: Establishing a grid system of standardized longitude and latitude coordinates for panoramic images; The standardized longitude and latitude coordinates of the established panoramic image are converted into the Hammer coordinate system to convert them into the x and y coordinates of the Hammer-Aitoff projection plane coordinate system. The calculation formula is as follows: ； in, represents the normalized latitude of the panoramic image, represents the normalized longitude of the panoramic image, z is the intermediate calculation quantity used to normalize x and y; Reprojecting the segmented image: After separating the RGB channels of the input segmented image, performing the Hammer coordinate system transformation on each color channel independently, and combining the transformed results of the three color channels to generate a projected segmented image; A corresponding category index is pre-assigned to each semantic segmentation element category, and a bidirectional mapping relationship between RGB color and category index is constructed, the projected segmented image is read, and the projected segmented image in RGB color space is converted pixel by pixel into the corresponding category index according to the pre-constructed bidirectional mapping relationship between RGB color and category index, so as to obtain a category index map for calculating the area of ​​each type of semantic segmentation element.

4. The public space quality assessment method based on multidimensional features according to claim 1 is characterized in that: The global visual feature index includes any one or more of a sky view factor, a building view factor, a green view factor, a visual entropy, a skyline fractal dimension, and a color property. The sky view factor uses the sky pixel ratio to measure the urban building density and sense of enclosure. The building view factor uses the building pixel ratio to measure the sense of enclosure and building density of the urban space. The green view factor uses the green vegetation pixel ratio to measure the degree of greening. The visual entropy is used to measure the complexity of information in the image. The skyline fractal dimension is used to characterize the spatial complexity of the building outline. The color properties include color brightness, contrast, and saturation. The local visual feature index includes any one or more of the proportion of natural elements, the proportion of traffic elements, the proportion of public facilities, the proportion of safety facilities, the proportion of human and vehicle activities, the proportion of street walls, the proportion of building signs, the proportion of street-facing windows, and the proportion of specified elements. The proportion of natural elements is used to characterize the proportion of natural elements in the image, the proportion of human and vehicle activities is used to characterize the proportion of the number of people and vehicles in the image, and the proportion of specified elements is the proportion of elements in the segmented image result that is greater than a preset proportion threshold or the proportion of elements of a specified type; The global spatial feature indicators include any one or more of the proportion of vertical elements, walking width, block length ratio, H / W ratio of street canyons, spatial connectivity and depth map spatial ratio. The vertical element proportion is used to characterize the relationship between building height and block scale, the walking width is the width of the street, the block length ratio is the ratio of block length and width, the H / W ratio of street canyons is the ratio of building height to street width, the spatial connectivity is used to characterize the traffic and connectivity between spaces, and the depth map spatial ratio is the ratio of different space types in the depth map. The panoramic depth map is generated by estimating the depth of the panoramic image using the unlabeled self-distillation v2 model, and the depth map spatial ratio is calculated.

5. The public space quality assessment method based on multidimensional features according to claim 1 is characterized in that: The use of a deep learning model to detect the category of public space functional facilities on the semantic segmentation processing results includes: establishing a facility text prompt set, inputting the semantic segmentation processing results into a Grounding DINO model for target detection and positioning; drawing a facility bounding box on the original panoramic image and marking the category label, and outputting the detection result with the marked information, wherein the marked information includes any one or more of the facility category, bounding box coordinates, detection confidence, and location information.

6. The method for evaluating the quality of public space based on multidimensional features according to any one of claims 1 to 5, characterized in that: The quality assessment index data set is clustered using a K-means clustering optimization algorithm based on an adaptive threshold. The public space types include mixed public spaces, open public spaces, and dense commercial blocks. The differences between the characteristic indicators of the mixed public spaces and the characteristic indicators of other space types are within a preset range. The sky view factor and green view factor of the open public space are higher than those of other space types and the excess reaches a first preset threshold. The building view factor of the dense commercial block is higher than that of other space types and the excess reaches a second preset threshold.

7. The method for evaluating the quality of public space based on multidimensional features according to claim 1, characterized in that: It also includes determining optimization strategies based on the evaluation results of the potential service capacity and actual attractiveness of each public space, including: Evaluate the actual attractiveness of the public space according to the actual flow of people in the public space to obtain an actual attractiveness evaluation result, wherein the actual flow of people data includes the number of visitors and residents; For the first area where the actual attraction of the public space is higher than the potential service capacity, the POI mix and the pedestrian level are optimized by adjusting the types of POIs and the number of pedestrian facilities. For the second area where the actual attraction of the public space is lower than the potential service capacity, the public services and visual effects are optimized according to the corresponding characteristic indicator values ​​in the quality evaluation indicator data set.

8. A computer device comprising a processor and a memory, wherein the memory is used to store a computer program, wherein: The processor is configured to execute the computer program to perform the method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed, the method according to any one of claims 1 to 7 is implemented.

Images