Data-Driven Approach to Assess Street Safety: Large-Scale Analysis of the Microscopic Design

Abstract

Safety is an important quality of street space that affects people’s psychological state and behavior in many ways. Previous large-scale assessment of street safety focuses more on social and physical factors and has less correlation with spatial design, especially the microscopic design. Limited by data and methods, street safety assessment related to microscopic design is mostly conducted on the small scale. Based on multisource big data, this study conducts a data-driven approach to assess the safety of street microscope design on a large scale from the perspective of individual perception. An assessment system including four dimensions of walkability, spatial enclosure, visual permeability, and vitality is constructed, which reflects the individual perceptions of the street space. Intraclass correlation coefficient (ICC) and location-based service (LBS) data are used to verify the effectiveness of the assessment method. The results show that multisource big data can effectively measure the physical elements and design features of streets, reflecting street users’ perception of vision, function, architecture, and street form, as well as the spatial selectivity based on their judgment of safety. The measurement of multidimensional connotations and the fusion of multiple data mining technologies promote the accuracy and effectiveness of the assessment method. Street safety presents the spatial distribution of high-value aggregation and low-value dispersion. Street safety is relatively low in areas with a large scale, lack of street interface, large amount of transit traffic, and high-density vegetation cover. The proposed method and the obtained results can be a reference for humanized street design and sustainable urban traffic planning and management.

1. Introduction

Safety, as one of the important qualities of street space [1], affects people’s psychological state and behavioral activities in many ways. A negative sense of safety reduces people’s willingness to engage in physical activities [2], while a positive sense of safety promotes the occurrence of social interactions [3] and increases residents’ satisfaction [4], sense of belonging [5], happiness [6], and cohesion [7], which in turn promotes their physical and mental health [8]. Safe streets have higher walkability [9], while unsafe streets reduce the quality of walking [10] and induce more motor vehicle travel [11,12]. In China, where there are more than 200,000 traffic accidents every year, motor vehicles cause 89.8% of them. Public security organs filed more than 4.5 million criminal cases, 70% of which took place on streets [13]. Moreover, in recent years, the European Union (EU) has placed great importance on road users’ safety (real and perceived), as it has considerable influence both in the operation of transport systems and on citizens’ quality of life [14,15]. The problem of street safety is becoming increasingly serious. It is of great significance for street safety design and public policy formulation to properly evaluate street safety and understand how spatial characteristics affect individual safety perceptions.

Although some studies have explored street safety assessment, several shortcomings still exist. First, large-scale assessment of street safety focuses more on social and physical factors and has less correlation with spatial design [16,17]. On the macro scale, street safety is an individual’s subjective perception and ability to feel safe from the street environment [18]. It is usually related to the fear of crime and other unsafe factors including crime, traffic, culture, individual characteristics, etc. [19]. Thus, large-scale assessment studies mostly measure street safety through safety incident data such as traffic accidents [15,20], crime rates [21], and robbery rates [22], which are more correlated with urban management, social security, and economic development. It is difficult to accurately assess the level of street safety design. In addition, some scholars question the accuracy of measuring street safety through criminal incidents, because the perception of street safety is empirical, and it is a rational judgment of residents’ fear degree caused by the street environment [19]. It is more related to environmental chaos, potential hidden dangers, and unknown fears, but not strictly related to actual criminal incidents [18].

Second, limited by data and methods, street safety assessment related to microscopic design is mostly conducted on the small scale. Street safety at the micro level is an individual’s visual perception and safety recognition of the physical environment, which emphasizes the connotations of spatial organization, urban imagery, sense of environmental affiliation, and social vitality [23,24,25]. Street users’ perceptions of spatial characteristics, including walkability limited by traffic conditions, enclosure created by scale and interface, permeability reflected by greenery and buildings, as well as vitality related to functions and density, can better explain the impact of safety design on pedestrians’ mental health and behavioral activities [24,26,27,28]. Although previous studies mostly use questionnaire surveys and on-site auditing methods to obtain basic data, higher time and labor costs make it difficult to carry out large-scale assessment [17,29,30]. In addition, questionnaire survey often carries subjective tendencies of the respondents, which leads to the incomparability of the assessment results between different regions. There are other studies using street view images for large-scale evaluation of microscope design [18,19]. As street view images are just human visual presentations, the results have limited direct guiding significance for street design [31,32].

Third, the previous safety assessment studies lack the consideration of the optimal values of assessment indicators. They consider that there is a simple linear relationship between the assessment result (e.g., safety level) and the indicator value [31] and ignore the possible nonlinear effects of these indicators. The ratio of distance to height of buildings along the street (D/H ratio), street length, and other indicators have optimal values. Street safety can be effectively improved only by controlling them within a reasonable range. In addition, the existing methods are mostly constructed based on classical theories and logical derivations, which lack effective tests based on realistic situations. This is also an important reason for the inconsistency of street safety assessment methods. People generally choose to move around in safe spaces, and the spatial choices of street users are an important reflection of the differences in street safety. At the macro level, the accuracy of the assessment methods can be effectively tested and calibrated by taking the spatial distribution of street activities as the standard.

Street view images, points of interest (POIs), building vector data, location-based service (LBS) data, and other multisource big data provide the possibility of a fine-grained perception of street space at the macroscopic scale. For instance, street view images are based on actual photos of streets taken from the pedestrian’s viewpoint. It can be quickly acquired over a large area at a low cost, which is conducive to perceiving the micro space from the macro scale. LBS data are the location points left by people operating mobile phone apps, which indicate that people are in a slow walking or static state. People will only slow down or stop when they feel that the external space of the street is safe, so LBS can reflect people’s psychological perception of street safety to a certain extent. With this psychological perception as the guide, the assessment method of street safety can be verified and calibrated. The abovementioned multisource big data provide an important data basis for individual perception-based street safety assessment on a larger scale.

By using multisource big data, this study constructs a data-driven approach to assess street safety related to microscopic design. The assessment system includes four dimensions: walkability, spatial enclosure, visual permeability, and vitality. The weights of 14 secondary indicators are calculated by the objective weighting method. After that, the independence of the assessment dimensions is verified by the intraclass correlation coefficient (ICC). The street users’ choice of safe streets is referred to by LBS data, and the street safety assessment system is tested. Thereby, the most effective street safety design assessment method is identified. Finally, this study takes Xiamen Island, the central urban area of Xiamen City, China, as a case, assesses its street safety, explores the spatial distribution and dimensional characteristics of street design safety at the macro level, and analyzes the reasons for the spatial heterogeneity.

This study is structured and organized as follows. Section 2 provides a literature review, including the knowledge structure of street safety research, the meaning of street safety related to microscopic design, and an overview of the assessment methods in previous studies. Section 3 introduces the data and methods. The data acquisition and preprocessing are described, and the selection of assessment dimensions and secondary indicators is discussed in detail. After that, the indicator weighting and the verification of the effectiveness of the assessment method are conducted. Section 4 is the case study of Xiamen Island. The spatial distribution of street safety related to microscopic design and the four dimensions are analyzed, and the causes of spatial heterogeneity are explored. Section 5 offers a discussion of this study, clarifying how the research questions have been answered, the comparison with previous studies, the challenges during the study process, and the possible deepening direction in the future. Section 6 is the conclusion, which is summarized from the data, methods, and results.

2. Related Work

2.1. Knowledge Structure of Street Safety Research

This study set the publication time from 2000 to 2022, entered the subject words “street safety”, “safe walking”, “pedestrian safety”, and “road safety”, and used the logical operator “OR” to search the literature in the core database of the Web of Science. Web of Science contains the world’s highest-level authoritative academic journals, monographs, and conferences, covering natural science, engineering technology, social science, and arts, etc. Its core database contains world-class academic research and browses the complete citation network. Book reviews, conference abstracts, book chapters, and literature that did not meet the professional research field were excluded. A total of 4076 papers were obtained. BiblioShiny software based on R language was used for bibliometric analysis.

2.1.1. Research Focus

To clarify the current research focus, we analyze the keyword frequency of the abstracts for the bibliometric analysis. Background words such as “street” and “road”, the search terms, and some keywords in sociology and traffic, including “road accident”, “traffic accident”, “crime”, etc., are excluded so as to avoid the deviation of the analysis results. Among the top 50 words, “road users” ranks first, indicating that the feeling of space users is the research focus (Figure 1). Words related to quantitative evaluation, such as “safety performance”, “safety measures”, “statistically significant”, “mobile phone” and “public health”, also appear frequently. The quantitative evaluation of street safety based on big data, which considers the feelings of street users, is an important research field at present.

2.1.2. Research Trend

In the early years, scholars paid more attention to motor traffic (Figure 2a). Research topics have become more focused through diversification since 2010. Then, we analyzed the themes of the publications after 2016 (Figure 2b). The first quadrant represents the important and well-developed themes (Motor Themes) [33], including pedestrian safety, crash data, and risk factors. The second quadrant represents highly developed and isolated themes (Niche Themes), including food safety, street food, safety performance, and rural roads. The third quadrant represents the emerging or declining themes, which have not developed well, including public health, developing countries, and cross-sectional study. The fourth quadrant represents the basic and transversal themes, which are currently popular, but not mature, and may become an important theme or research trend in the future, including statistically significant safety issues, road users, traffic safety, pedestrian safety, risk factors, and so on.

In summary, the knowledge structure of street safety shows the following characteristics: ① The quantitative research of street users’ safety is one of the current research focuses; ② The research direction has changed from motorized traffic to non-motorized and pedestrian traffic, especially pedestrian safety; ③ Pedestrian safety evaluation based on data and various factors is not only an important research direction but also a basic direction. Therefore, developing a data-driven approach to evaluate street users’ safety is consistent with the knowledge structure.

2.2. The Meaning of Street Safety

Safety has both objective and subjective meanings [34,35]. Objectively, safety refers to an individual’s freedom from threats and avoidance of dangers in the environment [34]. Subjectively, safety is an individual’s perceived judgment of environmental dangers [36]. Subjective safety is a more independent and important factor because people’s psychological perceptions of environments lead to differences in behavior in multiple ways [37,38,39,40,41,42]. This study explores the safety related to the microscopic design of streets. To further clarify the definition of street safety, we extensively collected studies on the safety of the built environment, summarized the theories related to street safety, and chose those with high citation rates and instructive significance. The related theories suggest that the physical elements provide a continuous sense of safety on the street (

Table 1. Theories of street safety.

Scholar	Year	Publication	Main Ideas	Morphological Elements
Jane Jacobs [43]	1961	The Death and Life of Great American Cities	The sense of street safety comes from the continuous pedestrian flow on the street and the informal “natural surveillance” provided by the stores.	Clear public and private boundaries, buildings and stores along the street, adequate and continuous pedestrian flow
Jan Gehl [44]	2003	Life between Buildings	Safe streets create social places where people can rely on and interact with each other.	Suitable walking space, places to stay, flexible boundaries
C. Ray Jeffery [45]	1971	Crime Prevention through Environmental Design	The built environment and facilities are designed to prevent and reduce crime, reduce pedestrian fear and concern, and increase sense of safety.	Form, openness, recognizability, and visibility of space, layout of buildings, and streetscape amenities
Oscar Newman [46]	1972	Defensible Space: Crime Prevention through Urban Design	The built environment is designed and adapted to reduce crime. The core idea is to increase informal surveillance and visibility of locations by enhancing visual accessibility.	Territoriality, surveillance, image, milieu
James Q Wilson; George L Kelling [47]	1982	Broken Windows	Disordered spaces and low-quality environments can induce potential criminal motivation and create negative safety experiences.	Well-maintained street space, neat layout of facilities, high-quality landscape greenery
Jay Appleton [48]	1984	Prospect Refuge Revisited	High visual permeability can promote individuals’ sense of self-protection in street space.	Transparent space and open view
Ito Zi [49]	1982	Urban Crime	The time blind spot caused by low monitoring coverage and the dead space created by the shading of buildings or obstacles are the main inducements of urban crime.	Clear spatial boundaries, permeable spaces, complete safety facilities
Peter Calthorpe [50]	1993	The Next American Metropolis: Ecology, Community & the American Dream	Based on the relationship between urban economic development, population health, green and low-carbon development and urban walkability, the author expounds on the importance of walkability to street safety.	Walkable space, convenient public transportation, away from the city’s main roads

).

Although street safety includes many dimensions such as society, traffic, and space. In this study, street safety refers to the safety recognition and perception of the streets’ microscopic design [43,44,45,46]. It is the ability of individuals to gain freedom and confidence and eliminate fear and anxiety from the physical environment [45,47,48]. As street safety is an individual’s judgment of the physical street environment, it is closely related to the morphological elements [43,45,49,50]. It emphasizes the connotations of spatial organization, urban image, functions, vitality, and sense of environmental affiliation [23,24,25]. These connotations are embodied in the walkability, enclosure, visual permeability, and vitality of streets [26,27,28,37,38].

2.3. Overview of Street Safety Assessment Methods

2.3.1. Traditional Methods of Street Safety Assessment

Questionnaires and systematic social observation (SSO) are usually used to obtain individuals’ perceptions of street safety [8]. The questionnaire is the most common method for assessing safety perceptions [7,51,52,53], pedestrian safety [25,26,54,55,56,57], crime-related safety [53], the perceived safety of specific populations [7,12,58], and influencing factors of safety perceptions [24,29,54,59]. The advantage is that it enables an in-depth analysis [60], which may be more interpretable than objective data [18,61]. The limitations are the high cost of manual processing and the time-consuming acquisition, which make it difficult to apply to large-scale and multi-sample studies.

SSO is a researcher-oriented subjective survey method. Researchers conduct field research in the study area, record subjective perceptions, and complete assessment reports [4,26,56,62]. Compared with the questionnaire method, the results of SSO are comparable across regions because they come from the same investigators; professionally trained investigators can uncover more comprehensive and in-depth information. However, the method is also time-consuming and costly because it requires investigators to traverse all study areas.

2.3.2. Street Safety Assessment Methods Based on New Data

To overcome the above limitations, new data have begun to be employed, including street view images, crowdsourced data, and building vector data. These multisource data are human-scale and have a wide range of time and space, which are free and easy to obtain for many cities and are more time-effective and cost-effective. The online mapping service platform has opened up access to street view images [8,32,63], which can be combined with the SSO [64] and machine learning [8,19,63,65,66,67] to predict street safety and score the street environment. The development of crowdsourcing technologies has allowed researchers to use web-based questionnaires as an alternative to on-site research [31,61,68]. The advantage of crowdsourcing data is that researchers do not need to travel to the field, increasing the efficiency of data acquisition. These data provide the possibility of assessing individuals’ perceptions of street safety on a large scale.

3. Data and Methods

3.1. Data

(1) Street network

Open Street Map (OSM) is widely used in the study of street accessibility [69], land use distribution [70], and spatial quality of streets [71]. We obtained street network data from OSM with 5637 entries. The original data were topologically checked and processed to remove expressways, elevated roads, and tunnels that lacked sidewalks. A total of 2466 residential streets were finally reserved. Referring to “Xiamen Urban Rail Transit Construction Plan (2011–2020)”, the streets are divided into four classes: expressways, arterial roads, secondary trunk roads, and branch roads.

(2) Street view image

Baidu Maps is the most widely used online map service platform in China. The street view images were obtained from the Baidu Street View API interface in June 2021 and were sampled at 40 m intervals. For streets that are too short in length, the centerline midpoint was selected. A total of 21,486 panoramic photos were crawled with a resolution of 4096 × 1380. The bottom third of the original image is occupied by the capture vehicle, so the bottom third of all images was cropped in batches. The elements in the street view images were identified by using the Deeplap V3+ semantic segmentation tool and Cityscapes Dataset. The identification results were divided into 5 categories and 19 semantic labels, including buildings, sky, plants, vehicles, roads, and other elements.

(3) POI

The POI data were obtained from the AutoNavi Open Platform in June 2021, containing 169,164 records and totaling 21 major categories and 236 minor categories. With reference to previous research, the POIs were reclassified into 8 categories: residential (2.86%), business (14.21%), government agencies and social groups (1.87%), commercial (66.37%), education (4.65%), green space (0.33%), transportation (5.23%), and others (4.47%).

(4) Building

The building data were obtained from open-source data provided by the Baidu Map open platform in 2021. The location and number of floors were checked based on QGIS online maps and street view images. The function was obtained from the current land use data provided by the Xiamen government website and checked by street view images. After preprocessing, there are 49,558 pieces of data.

(5) House price

The house price data were obtained from the sale information released by the Lianjia property transaction website in 2021. Lianjia (http://www.lianjia.com/, accessed on 24 June 2021) is one of the major full-chain property service platforms in China with comprehensive information. We obtained a total of 1031 houses on Xiamen Island, including the building age. On this basis, we manually supplemented the building age of 87 major public buildings and commercial center buildings to obtain relatively complete building age data.

(6) LBS data

Mobile internet location services actively initiate or passively generate LBS data, recording accurate GPS-accurate locations (up to 10 m) and corresponding time stamps. The high accuracy makes LBS data especially suitable for street research. A trajectory represents a mobile phone user slow walking or stopping at that location. The LBS data are anonymous geographic data provided by Aurora Mobile. Aurora Mobile provides an SDK location development environment for smartphone applications and monitors approximately 1,452,000 mobile application terminals. Data collection mechanisms include regular updates and event triggers.

The data were collected from 17 to 30 October 2020, including 10 weekdays and 4 weekends with no holidays. During this period, the impact of COVID-19 on Xiamen diminished. The temperature was suitable (19.8–26.9 degrees Celsius), and the weather was rain-free. The original data were processed as follows: ① All of the LBS data present in Xiamen from 17 to 30 October 2020 were selected, containing 130,817,778 records and 6.021 million users. The daily data volume is relatively stable; ② The users’ location information in the streets from 8:00 to 22:00 every day was calculated by advancing one hour. Up to 329,800 users were identified on weekdays (19:00 on 19 October) and up to 309,200 users on weekends (18:00 on 17 October), while there are approximately 2.1 million permanent residents on Xiamen Island; ③ The multiday averages of weekdays and weekends were summarized, respectively.

3.2. Assessment Dimensions and Indicators

A street is a composite space that carries residents’ daily life, social activities, and transportation functions [16,43]. Street space is defined as the space surrounded by buildings on both sides of the street. The assessment framework of street safety includes four dimensions: walkability, spatial enclosure, visual permeability, and vitality.

(1) Walkability

Walkability is the fundamental property of street safety that reflects how pedestrian-friendly a street is [72]. “Walkability” is defined as the ability to encourage walking by providing a comfortable and safe walking environment for pedestrians [73]. People are more likely to think that a pedestrian-friendly street environment is safe [57]. Potential traffic hazards such as lack of traffic safety facilities [59], narrow sidewalks, and higher traffic density, speed [74], and intersection density can reduce pedestrians’ perception of street safety. Therefore, four indicators are selected to assess walkability: the relative walking width index, vehicle interference index, traffic facilities index, and pedestrian appearance ratio.

(2) Spatial enclosure

Spatial enclosure is the sense of enclosure produced by the physical elements that define the space boundary. Spatial enclosure gives the street an identifiable internal space, making the street not only a medium for carrying traffic functions but also a place with usage significance [75]. A sense of enclosure can enhance people’s identification of their territory through clear boundaries, creating a sense of “territoriality” to the street [76], and thus enhancing the perception of safety. Physical elements that enclose street space include buildings [43,77], walls [78], fences [79], trees [80], etc. The proportional relationship between street width and building height on both sides is considered to be closely related to the actual experience of pedestrians [81]. Therefore, this study chooses three indicators to assess the spatial enclosure: distance to the optimal D/H ratio, distance to the optimal interface continuity index, and street view enclosure index.

(3) Visual permeability

Visual permeability reflects the degree of an individual’s sight occlusion in the space. There is no logical conflict between visual permeability and spatial enclosure. The street interface with visible shop windows and the orderly layout of the landscape can also have a transparent vision while maintaining a high enclosure. People will spontaneously choose open environments with good vision, which helps to detect potential safety hazards and stimulates their awareness of self-protection in public spaces [45,46,48]. Trees are the main elements affecting visual permeability [19,82]. The random arrangement of facilities can block the pedestrian’s sight. By improving the possibility of visual interaction between street space and building interiors, the opportunities for people to come and go will increase [83], which creates potential surveillance. Therefore, the visual obscuration index, interface transparency index, and sky openness index are chosen to assess the visual permeability of streets.

(4) Vitality

Safe streets can attract more residents’ daily activities than unsafe streets. The theory that street vitality affects the perception of street safety is based on Jacobs’ “street eye” theory [43] and Newman’s defensible space theory [46]. Street safety comes from an interconnected informal social network, which is generated by pedestrians and retailers. It plays a role in monitoring and inhibiting criminal behavior, which is called the “natural surveillance” of the street. Diversity, density, small street segments, and older buildings work together to stimulate street vitality [43]. Therefore, functional diversity, development intensity, distance to the optimal street length, and building age are used to assess street vitality [84,85].

The description and calculation methods of the assessment indicators of the above four dimensions are shown in

Table 2. The description and calculation method of street safety assessment indices.

Dimension	Indicator	Description of Indicator	Formula	Explanation	Data Type
Walkability	Relative Walking Width Index	The ratio of walkway to all traffic space, reflecting the walking capacity of the street	$RWWI=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_walk}}{S_{i\_\mathrm{road}}+S_{i\_vehicle}+S_{i\_walk}}}$	$RWWI$ is the relative walking width index; $S_{i\_walk}$, $S_{i\_\mathrm{road}}$, $S_{i\_vehicle}$, respectively, represent the pixel area of the walkway, roadway, and vehicles in the i-th street view image of the street. m is the number of street view images collected for the street.	Street View Image
	Vehicle Interference Index	The ratio of vehicles to roadway, reflecting the degree of disturbance to pedestrians by vehicles	$VII=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_vehicle}}{S_{i\_\mathrm{road}}+S_{i\_vehicle}}}$	$VII$ is the vehicle interference index, $S_{i\_\mathrm{road}}$, $S_{i\_vehicle}$, respectively, represent the pixel area of roadway and vehicle in the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
	Traffic Facilities Index	The ratio of traffic safety facilities to the street view image, reflecting the traffic management level of the street	$TFI=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_fac}}{S_{i\_SV}}}$	$TFI$ is the traffic facility index, $S_{i\_fac}$ is the pixel area of traffic signals and signs in the i-th street view image of the street, $S_{i\_SV}$ is the total pixel area of the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
	Pedestrian Appearance Ratio	The ratio of pedestrians to the street view image, reflecting the density of pedestrians on the street	$PAR=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_ped}}{S_{i\_SV}}}$	$PAR$ is the pedestrian appearance rate, $S_{i\_ped}$ is the pixel area of pedestrians in the i-th street view image of the street, $S_{i\_SV}$ is the total pixel area of the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
Spatial Enclosure	Distance to the Optimal D/H Ratio	The optimal D/H ratio is subtracted from the ratio of the average street width to the average building height, and then the absolute value is calculated. The optimal D/H ratio is determined as 1.5 [86,87].	$D/H_{D\mathrm{ist}\_OV}=\left|\frac{D_{Ave}}{H_{Ave}}-1.5\right|$	$D/H_{D\mathrm{ist}\_OV}$ is the distance to the optimal D/H ratio, $D_{A\mathrm{ve}}$ is the average street width, and $H_{A\mathrm{ve}}$ is the average height of the buildings along the street.	Building vector data
	Distance to the Optimal Interface Continuity Index	The optimal build-to-line ratio is subtracted from the average build-to-line ratio, and then the absolute value is calculated. Combining the construction experience of Western countries and Chinese major cities, the optimal build-to-line ratio is determined as 0.8 [88]. The calculation refers to the quantitative identification method of street interface based on the GIS platform proposed by Harvey [89].	$IC{I}_{Dist\_OV}=\left|\frac{l_{\mathrm{sum}}}{L}-0.8\right|$	$IC{I}_{D\mathrm{ist}\_OV}$ is the distance to the optimal interface continuity index, $l_{\mathrm{sum}}$ is the total length of the building interface along the street, and $L$ is the length of the street centerline.	Building vector data
	Street View Enclosure Index	The ratio of buildings, walls, fences, and other elements that have the function of defining the spatial boundary to the street view image	$SVEI=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_bui}+S_{i\_wall}+S_{i\_fen}}{S_{i\_SV}}}$	$SVEI$ is the street view enclosure index,$S_{i\_bui}$, $S_{i\_wall}$, $S_{i\_fen}$, respectively, represent the pixel area of buildings, walls, and fences in the i-th street view image of the street, $S_{i\_SV}$ is the total pixel area of the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
Visual Permeability	Visual Obscuration Index	The ratio of trees, pillars, vehicles, and other elements that obstruct the sight to the street view image, reflecting the degree to which the environmental elements in the street space obstruct the sight of pedestrians.	$VOI=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_tree}+S_{i\_pole}+S_{i\_vehicle}}{S_{i\_SV}}}$	$VOI$ is the visual obscuration index, $S_{i\_tree}$, $S_{i\_\mathrm{pole}}$ $S_{i\_vehicle}$, respectively, represent the pixel area of trees, pillars, and various vehicles in the i-th street view image of the street, $S_{i\_SV}$ is the total pixel area of the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
	Interface Transparency Index	The ratio of the horizontal length of the building interface with visual permeability to the total length of the building interface along the street. Building interfaces are classified into four categories: commercial building interfaces with high permeability (category I), office building interfaces with medium permeability (category II), residential building interfaces with low permeability (category III), and walls without permeability (category IV) [90].	$ITI=\frac{l_I\times 1.25+l_{II}+l_{III}\times 0.75}{l_{\mathrm{sum}}}$	$ITI$ is the interface transparency index, $l_I$ is the length of the category I building interface of the street, $l_{II}$ is the length of the category II building interface of the street, $l_{III}$ is the length of the category III building interface of the street, and $l_{\mathrm{sum}}$ is the total length of the building interface along the street.	Building vector data
	Sky Openness Index	The ratio of the sky to the street view image, reflecting the sky openness of the street space	$SOI=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{S_{i\_sky}}{S_{i\_SV}}}$	$SOI$ is the sky openness index, $S_{i\_sky}$ is the pixel area of the sky in the i-th street view image of the street, $S_{i\_SV}$ is the total pixel area of the i-th street view image of the street, and m is the number of street view images collected for the street.	Street View Image
Vitality	Functional Diversity	The mix of POIs within 55 m around the street, reflecting the diversity of street functions. The POIs are classified into eight types: government institution, transportation, commerce, education, housing, company and enterprise, green, and others.	$Diversity={\displaystyle \sum _{r=1}^n\frac{P_r\ast \ln \left(P_r\right)}{\ln \left(\mathrm{n}\right)}}$	$Diversity$ is the functional diversity around the street, $P_r$ is the proportion of r-th POI to the total number of POI within 55 m around the street, and n is the number of POI types within 55 m meters around the street.	POIs
	Development Intensity	Building floor area ratio within 100 m around the street	$FAR=\frac{S_{bui}}{S_{land}}$	$FAR$ is the development intensity around the street, $S_{bui}$ is the total building area within 100 m around the street, and $S_{land}$ is the total land area within 100 m around the street.	Building vector data
	Distance to the Optimal Street Length	The optimal street length is subtracted from the actual street length, and then the absolute value is calculated. The optimal street length is determined as 100 m [91,92,93].	$L_{D\mathrm{ist}\_OV}=\left|l-100\right|$	$L_{Dist\_OV}$ is the distance to the optimal street length, and $l$ is the length of the street centerline.	Open Street Map
	Building Age	The age of the buildings around the street	$A\mathrm{ge}=2022-N_{bui}$	$Age$ is the building age around the street, and $N_{bui}$ is the completion year of the nearest residential neighborhood or public building around the street.	Building vector data

.

3.3. Indicator Weighting

The entropy method can eliminate the human subjective factor in weight determination and has been widely used in street assessment research [94]. The smaller the entropy value, the higher the dispersion of the indicator, which means the greater the influence (weight) of the indicator on the comprehensive evaluation [95]. The calculation steps are as follows:

(1) Dimensionless processing. Since the dimension and positive and negative orientation of each indicator are different, the original data need to be formalized. The formulas are as follows:

$$\mathrm{Positive} \mathrm{indicator}: X_{ij}^{\prime }=\frac{x_{ij}-\min \left(x_j\right)}{\max \left(x_j\right)-\min \left(x_j\right)}$$

(1)

$$\mathrm{Negative} \mathrm{indicator}: X_{ij}^{\prime }=\frac{\max \left(x_j\right)-x_{ij}}{\max \left(x_j\right)-\min \left(x_j\right)}$$

(2)

where $X_{ij}^{\prime }$ is the dimensionless processed result of the j-th indicator of the i-th street,$x_{ij}$ is the value of the j-th indicator of the i-th street, and $\max \left(x_j\right)$ and $\min \left(x_j\right)$ represent the maximum and minimum values of the j-th indicator, respectively.

(2) Entropy value calculation:

$$E_j=-k{\displaystyle {\sum }_{i=1}^m\left(Y_{ij}\times \ln Y_{ij}\right)}$$

(3)

$$\mathrm{k}=\frac{1}{\ln m}$$

(4)

$$Y_{\mathrm{ij}}=\frac{X_{ij}^{\prime }}{{\displaystyle {\sum }_{i=1}^m{X}_{ij}^{\prime }}}$$

(5)

where $E_j$ is the information entropy value of the j-th indicator and $m$ is the sample size.

(3) Weight calculation:

$$W_j=\frac{d_j}{{\displaystyle {\sum }_{j=1}^n{d}_j}}$$

(6)

$$d_j=1-E_j$$

(7)

where $W_j$ is the weight of the j-th indicator and $n$ is the number of indicators.

By assigning weights to the assessment indicators of the four dimensions, the assessment system of street safety is constructed. Next, the validity of the assessment system needs to be tested.

3.4. Validity Test of the Assessment System

(1) Independence test of assessment dimensions

Since the indicators are all continuous variables, the intraclass correlation coefficient (ICC) is used to conduct a consistency test to check whether the four dimensions are independent of each other. This is an important foundation before determining the assessment system. The two-way random model is chosen to improve the broad applicability of the test results and to ignore the possible systematic errors [96]. The ICC assessment criterion that has been generally accepted is ICC < 0.4, indicating low consistency between observed data, and ICC ≥ 0.75, indicating high consistency [97]. The formula is as follows:

$$\mathrm{ICC}=\frac{BMS-WMS}{BMS+\left(k-1\right)WMS}$$

(8)

where $\mathrm{ICC}$ is the intraclass correlation coefficient, $BMS$ is the mean variance of subjects, $WMS$ is the mean variance of random errors, and $k$ is the number of measurement repetitions.

(2) Validity check based on the spatial distribution of residents

The spatial distribution of people in the street is used as a criterion to test the validity of the method. If the two have a certain correlation, it shows that the street safety assessment method constructed in this study is reasonable. It has become quite common for urban residents to use mobile phones for reading, socializing, navigation, shopping, and other mobile internet services in their daily lives. On the streets, people generally choose to use their mobile phones in a space where they feel safer, and almost no one will stop operating the mobile app in a location with heavy traffic or pedestrian flow (Figure 3). Therefore, LBS data can reflect people’s subconscious choices to a certain extent.

The frequency of people’s walking slowly or stopping in the street space, which means activity intensity, is used as a test criterion. The calculation method is as follows:

$$Intensit{y}_i=({\displaystyle \sum _1^t\frac{LBS\_da{y}_{i,t}}{Are{a}_i}}\times 5+{\displaystyle \sum _1^t\frac{LBS\_en{d}_{i,t}}{Are{a}_i}}\times 2)/7$$

(9)

where $Intensit{y}_i$ is the activity intensity of the i-th street (person/m²); $LBS\_da{y}_{i,t}$ is the number of active LBS records in the i-th street in the daytime on a weekday, $LBS\_en{d}_{i,t}$ is the number of active LBS records in the i-th street in the daytime on weekends, and both are multiday averages; $Are{a}_i$ is the area of the street; and $t=8,9,\mathrm{...22}$.

The correlation between the activity intensity of people moving slowly and stopping in the street and the assessment results of street safety is calculated, thus testing the validity of the street safety assessment method. Figure 4 summarizes the workflow.

4. Results

4.1. Study Area

This study presents a quantitative assessment of street safety on Xiamen Island. Xiamen is the central city of the southeastern coastal region of China and the head city of the western strait urban agglomeration. The subtropical climate and unique island-type landscape make it one of the most livable cities in China. The city has evolved through the history of tenancy, sea defense, and the establishment of special economic zones, resulting in a diverse street system and unique street space forms (Figure 5). Xiamen Island is the central urban area of Xiamen, the origin of the city, and an important political, cultural, educational, and commercial center. Xiamen Island contains two administrative districts. In 2021, the city had approximately 2.1 million permanent residents and a 132.2 km² urban built-up area, with high population density, development intensity, street network density, and convenient service facilities. There are a variety of street space forms and block forms, which makes it a typical sample for conducting street safety studies.

4.2. Validity Test Results

The weight calculation results of the entropy method are shown in Figure 6. Walkability has the highest weight among the four dimensions and has the highest contribution to the assessment results. The weights of spatial enclosure, visual permeability, and vitality are similar. Based on the weight calculation results, the independence of the four dimensions was tested. There is no strong correlation between them (Figure 7). The ICC value is 0.018, and the four dimensions have high independence from each other. Finally, the correlation between activity intensity and street safety was calculated. The correlation coefficient is 0.542, which is significant at the level of 0.001, indicating that the street safety assessment method constructed in this study has strong rationality and validity.

4.3. Spatial Distribution of Street Safety

The spatial distribution of street safety on Xiamen Island is high-value clustering and low-value dispersion. In general, the assessment values are higher in the southwest area, lower in the northeast area, higher in the core area of the island, and lower in the edge areas (Figure 8).

The streets with higher safety perceptions are mainly located on Amoi YatSen Road, south of Yundang Lake and along Xiahe Road, with a small number in the area north of Xianyue Park and Guanyinshan Business Center. The streets with lower safety perceptions are mainly distributed in the Gaoqi community, the airport area, and south of Wuyuan Bay. The typical streets in the area are selected for further analysis, and we found that compared with the streets with lower safety, the streets with higher safety generally have the following characteristics: clear spatial boundaries, neat and continuous interface, abundant street shops, small-scale space, and neat street space, while the streets with lower safety are generally socially and economically backward, and public security in the area is low. These streets lack safety design and management, so the street space is characterized by large-scale space, lack of street wall enclosure, a large amount of transit traffic, high vegetation density, and sparse pedestrian flow.

4.4. Dimensional Characteristics of Street Safety

(1) Walkability

The assessment result of walkability on Xiamen Island does not show significant regional differences (Figure 9a), but it is significantly correlated with road level. Walkability decreases as road level increases (using the median to reflect the general level of walkability at different levels, 0.0627 for branch roads, 0.0481 for secondary trunk roads, and 0.0335 for arterial roads and expressways). The main reason is that the street level in Chinese cities is generally classified by the volume of motor vehicles, and high-grade streets often carry heavy traffic. The safety hazards and the noise and exhaust from motor vehicles reduce walkability. The functions of low-grade streets are mainly for living services. The street space is designed to reduce traffic speed, and there is more pedestrian space, which reduces the interference of motor vehicles with walking comfort and improves walkability.

(2) Spatial enclosure

The assessment result of the spatial enclosures on Xiamen Island shows the spatial distribution characteristic as high in the southwest and low in the northeast, and the high values show single-core concentration and multipoint dispersion (Figure 9b). The high values are concentrated in the old urban area in the southwest, which are historic districts. The buildings are mainly low and middle-rise with 3–4 floors, and the street width is 10–15 m, forming a scale suitable for various activities. The cavalier-style buildings along the street form a neat and continuous interface, forming highly ordered space and enhancing the visual continuity of pedestrians. The high D/H ratio strengthens the streets’ spatial boundaries and enhances the sense of spatial enclosure. The low values are scattered in the outer areas in the northeast, where the development intensity is low and the streets are empty and sparsely built on both sides, resulting in a lack of spatial enclosure. It is found that buildings are the core element affecting the spatial enclosure of streets. The height of buildings along the streets, the interval distance between buildings, the neatness of building arrangement, and the building density affect the sense of spatial enclosure (Figure 10 and Figure 11).

(3) Visual permeability

The assessment results of the visual permeability show a hollow spatial distribution with high values in central areas and low values in fringe areas (Figure 9c). The streets with high visual permeability are mostly landscape streets located in the new urban areas on the fringe of Xiamen Island, which generally have high spatial quality and a neat layout of street facilities. The streets are wide and straight in the form of a square road network, and street space has high visual continuity and low visual interference. In addition, such areas are commercially developed, with numerous street shops and permeable street interfaces. The visual permeability of the streets in the central area of Xiamen Island is lower because there are a large number of old and closed communities, with a narrow scale and loose security management. This leads to a high density of street vegetation, disorderly layout of facilities, random parking of motor vehicles, and illegal construction of shops along the street. These streets have a large number of visually blind areas, the enclosed walls reduce the interface permeability, and thus, the spatial visibility is relatively low.

(4) Vitality

The assessment results of street vitality on Xiamen Island show a high spatial distribution in the southwest and low in the northeast (Figure 9d). The overall structural unfolding degree is insufficient, and the assessment result shows the characteristics of small clustering and large scattering. The southwest areas are the earliest development and most mature core areas in Xiamen, with superior locations and good landscapes. Aggregated resources, high development intensity, and well-preserved old buildings bring continuous and dense pedestrian flow to the streets. The street scale is suitable, and the average street length is 100–200 m, which is suitable for people’s short-distance travel. In addition, the open blocks enhance the interaction between people and space and create diverse street shops, which promote vitality. The street vitality in the northeastern fringe of Xiamen Island is generally low; the main reason is that there are many large areas with a single function, such as airports and industries. The streets are large in scale and lack street shops and closed interfaces, which leads to low pedestrian willingness and causes a continuous decline in street vitality.

5. Discussion

Different from the previous large-scale measurement of social factors such as traffic accidents [16,19] and small-scale evaluation of micro-design [23,24,25], this study aims to develop a data-driven approach to assess the safety of street microscope design on the large scale from the perspective of individual perceptions. Multisource big data and data mining are used to construct an assessment system including the four dimensions of walkability, spatial enclosure, visual permeability, and vitality, which are emphasized in the existing literature [43,45,46,48,72,75]. Based on this evaluation method, the heterogeneous distribution of street safety at the macro level can be accurately described, and the unsafe areas can be optimized by analyzing their influencing factors, so as to improve spatial safety and land use efficiency. Results indicate that street safety shows the characteristics of spatial agglomeration, and the large scale, lack of street walls, and high-density vegetation lead to a low level of street safety. The results also manifest that the data-driven approach has strong effectiveness and accuracy in large-scale assessment.

Safety has both objective and subjective meanings [35]. Although there is a certain difference between “perceptions of safety” and actual safety, this study explores the safety related to the microscopic design of streets. If safety is put in the context of street micro-design, the evaluation of street safety focuses more on street users’ judgment of space safety [37,41]. Microscopic design can influence people’s “safety perception” more. This study also considers actual safety. LBS data are used to refer to actual safety and verify the effectiveness of the street safety evaluation method. To some extent, the exploration of street safety in this study is more of a comprehensive analysis.

The definition of street safety in this study comes from classical theories [43,44,45,46,47,48,49,50]. The difference is that we use new data and methods to realize the large-scale and accurate measurement of street safety design, which is difficult to achieve by traditional methods. The evaluation of street safety and the validation of the approach in this study are all based on individual perception data. This is conducive to improving the individual’s space experience in the street [63,64]. Moreover, the optimal values are employed to calculate the indices, which breaks through the simple assumption of a linear relationship between environmental elements and street safety.

Although LBS data can reflect pedestrians’ subconscious judgments about street safety to a certain extent [85], the data can only be used for validity tests and cannot be used directly for street safety assessment. First, LBS data are not directly collected with street design orientation in mind but are generated when people operate mobile phone apps [98]. As a kind of big data, although LBS data contain a large number of samples, they are still affected by sampling bias and cannot fully and accurately reflect the psychological perceptions of all street users [85]. Second, LBS data cannot reflect the refined characteristics of street safety. For example, quantitative indicators such as the D/H ratio and interface transparency degree cannot be mined in LBS data, which makes it difficult to analyze the causes of the spatial heterogeneity of street safety. Nevertheless, LBS records in outdoor spaces are still a referential judgment for safe streets.

Although both spatial enclosure and visual permeability are related to the interface, there is no logical conflict between the two. They describe different feelings of individuals in the street. Visual permeability reflects spatial visibility and the ability to observe the environment [45,48]. Visual blindness caused by obstacles is an important factor that induces safety problems [49]. Spatial enclosure emphasizes the influence of places enclosed by elements such as buildings, walls, trees, and scale on safety perception [43,77,78,79,80]. The street interface with visible shop windows and a well-organized landscape also has a transparent view on the basis of maintaining good enclosure [43,45]. Combining visual permeability with spatial enclosure can explain street safety dialectically.

The street level has a certain influence on the assessment of street safety. We mainly deal with this influence from the following two aspects. First, we set the indicators of vehicle interference and traffic facilities, which can reflect the information on the street level. Second, when calculating other indicators, such as functional diversity and development intensity, we set different buffer ranges according to their implications and calculation requirements. We assume that the differences in street level can be reflected in the built environment characteristics of streets, and the buffer ranges reflect the general feelings of street users and the land use features [98].

The rapid development of technology can provide more deepening directions for related research in the future. The popularization of virtual reality (VR) technology makes it possible to study immersive feelings in laboratory conditions, as well as the scientific acquisition and quantitative analysis of subjective perception and street safety. Wearable physiological sensor technology, especially the increasingly mature integration of EEG sensors and skin sensors, provides a new way to objectively measure the feelings of individuals. Although the above technologies are mostly used in the research of meso and micro scales at present, combined with urban 3D models, it is expected to break through the limitation of abstract experience perception in previous studies. In addition, how to formulate effective measures to improve safety in low-value areas is a direction to be explored in the future. Instead of constructing a model to explore which factors affect street safety and predict it, an evaluation method is developed. How to intervene in the indicators to improve street safety is the next research direction. Street walking facilities can be increased to promote spatial interaction and spontaneous supervision. Through gradual street renewal, the spatial scale and the functional diversity should be controlled in reasonable ranges, and the old buildings should be protected and used in an orderly manner. City managers can formulate detailed rules to improve street safety.

6. Conclusions

This study implemented a data-driven approach to assess street safety on the large scale. It is found that multisource big data, including street view images, POIs, building vector data, LBS data, housing price, etc., provide the possibility to evaluate individual perceptions of street design at the macro level. Compared with previous studies that used a single data source to evaluate street safety, multisource data can measure the physical elements and design features more comprehensively. The data set in this study reflects street users’ perception of vision, function, architecture, and street form, as well as the spatial selectivity based on their judgment of safety, although these perceptions are sometimes realized through mobile network technology. This kind of perception is the street users’ understanding and feedback of the microscope design, which can be used to evaluate the safety of street design.

In terms of methods, the assessment of multidimensional connotations and the fusion of multiple data mining technologies make the results more accurate. First, the integration of deep learning, image semantic segmentation, and GIS spatial analysis can effectively extract street design elements from the data set. The optimal values of the assessment indicators are taken into consideration to avoid possible errors and made the assessment more reasonable and operable. Second, the four dimensions of walkability, spatial enclosure, visual permeability, and vitality can assist in large-scale quantitative display and demonstration of street safety issues. Third, the verification of the assessment method based on ICC and LBS data is necessary, which can reduce the deviation of the results. The verification results confirm the accuracy and effectiveness of the assessment method.

The results show that street safety presents the spatial distribution of high-value aggregation and low-value dispersion. Street safety is higher in core areas and lower in fringe areas. Old urban areas and business districts have higher street safety. Street safety is low in areas with a large scale, lack of street interface, large amount of transit traffic, and high-density vegetation. Street walkability is significantly related to the street level and decreases with the upgrading of the street level. The high-value areas of spatial enclosure show the characteristics of single-core concentration and multipoint dispersion, and the building form is the core affecting element. Streets with high visual permeability are mainly distributed in landscape areas and new urban areas. Streets in areas with aggregated resources and high development intensity have high vitality.

The data-driven approach proposed in this study provides a decision-making reference for the “humanized” design of streets and sustainable urban transportation planning and management. Rapid automated assessment methods and techniques based on big data can be used to conduct large-scale assessments at low cost. This approach is conducive to achieving an accurate assessment of individual perceptions and can be applied to other studies that require data fusion and mining. These results contribute to the validation of interdisciplinary approaches that combine the strengths of computer science and statistics with environmental psychology and social sciences.