Investigating urban perception using procedural street generation and virtual reality

(1)

STOCKHOLM SWEDEN 2016,

Investigating urban perception

using procedural street generation and virtual reality

OSCAR FRIBERG

KTH ROYAL INSTITUTE OF TECHNOLOGY

(2)

street generation and virtual reality

Undersökning av den urbana perceptionen med hjälp av procedural generation av gatuscener och virtual reality

OSCAR FRIBERG OFRI@KTH.SE

Master’s Thesis in Media Technology

Master of Science in Engineering - Media Technology Royal Institute of Technology

Supervisor: Christopher Peters Examiner: Tino Weinkauf

Date: 04-07-2016

(3)

(4)

As virtual reality and interaction possibilities in game technology evolve and become more user-friendly, accessible and cheaper, this technology suddenly gets useful in occupational areas that traditionally have not used them, one being urban planning. By using a game engine and a virtual neighborhood that utilizes procedural generation the user has the option of having a cost-effective, replicable environment with great control from the user.

According to the Broken Windows Theory small differences in appearance can have a big impact on how an urban neighborhood will be perceived. Therefore it would be interesting to investigate how the perception of street scenes is affected by the different aesthetic properties of dirt, graffiti, broken windows, building height and greenery (trees and bushes). By implementing a virtual neighborhood that utilizes procedural generation the basic structure of the neighborhood can stay the same, while separately adjusting the specific properties that are being evaluated.

Through surveys and interviews the results show that every one of the chosen properties are very significant except for building height. Generally the more trees that were present, the more safe the neighborhood felt, and the order of significance of dirt, graffiti and broken windows differed between the two evaluation methods.

Sammanfattning

Då virtual reality och interaktionsmöjligheterna inom spelteknologi utvecklas och blir mer användarvänliga, mer tillgängliga och billigare, blir denna teknik plöt- sligt användbar inom yrkesområden som traditionellt sett inte har använt dem, där ett exempel är stadsplanering. Genom att använda en spelmotor och en virtuell stadsdel som utnyttjar procedural generation får användaren en kostnadseffektiv, replikerbar miljö med stor kontroll från användaren.

Enligt Broken Windows Theory så kan små skillnader i utseende ha en stor in- verkan på hur en stadsdel kommer att uppfattas. Därför vore det intressant att un- dersöka hur synen på virtuella gatscener påverkas av olika estetiska egenskaper som smuts, graffiti, trasiga fönster, bygghöjd, och grönska (träd och buskar). Genom att implementera en virtuell stadsdel som utnyttjar procedural generation kan grund- strukturen vara densamma, medan man separat justerar de specifika egenskaperna som ska utvärderas. Genom enkäter och intervjuer visar resultaten att var och en av de valda egenskaperna är mycket betydande för den urbana perceptionen, med undantaget för byggnadshöjd. Generellt så kändes stadsdelen mer säker desto mer träd som fanns, och ordningen av signifikansen för smuts, graffiti, och trasiga fönster skiljde sig mellan enkäterna och intervjuerna.

(5)

1 Introduction 1

1.1 Aim . . . 2

1.2 Objective . . . 2

1.3 Delimitations . . . 2

1.4 Methodology . . . 2

1.5 Structure . . . 3

2 Background 5 2.1 Procedural Generation . . . 5

2.2 Urban Perception . . . 7

2.2.1 Place-Pulse . . . 8

2.2.2 Broken Windows Theory . . . 9

2.2.3 Greenery . . . 10

3 Implementation 13 3.1 3D Modelling . . . 13

3.2 Texturing . . . 15

3.3 Procedural Generation . . . 18

3.4 Virtual Reality . . . 19

4 Evaluation 21 4.1 Evaluation Methodology . . . 21

4.1.1 Survey Implementation . . . 21

4.1.2 VR Test and Interviews Implementation . . . 22

4.2 Survey Results . . . 24

4.3 Virtual Reality Interviews . . . 28

4.4 Analysis and Discussion . . . 30

5 Conclusion 33 5.1 Future Work . . . 34

6 Acknowledgments 35

References 37

(6)

A Survey 39

(7)

(8)

Introduction

Small differences in appearance can have a big impact on how an urban area will be perceived, as conceived in the Broken Window Theory (covered in section 2.2.2).

For example minor damages such as a broken window on a building that is left un- fixed will increase the risk of more broken windows, which in turn increases the risk of people breaking into the building from where the disorder continues to escalate.

There are many different aesthetic qualities - such as building height, breakage and amount of greenery - that may or may not have an impact on how the neighborhood will be perceived in regards to safety.

Therefore it would be interesting to investigate what effect these different qualities of the neighborhood have on how an area will be perceived. Using game technology and implementing a virtual neighborhood that utilizes procedural generation, the user get a replicable environment which would not be possible in real life as the neighborhood would be less controlled, whether it being lighting, humans or weather. It also opens up the possibility of real-time modifications for density and intensity of properties such as greenery, buildings and breakage/damage. Us- ing a game engine the users can do a walkabout in the virtual environment for themselves, that also have the possibility of being somewhat unique for each time generated while still maintaining the same main structure. Using virtual reality headsets the experience becomes even more realistic and immersive, as the user can place themselves in the scene that could relatively realistically mimic being in a real life neighborhood, looking around using head movements as well as getting a three-dimensional perspective.

To narrow it down I will be focusing on the building aspect of the area where parameters will be building height, dirtiness, graffiti, window breakage, and greenery (bushes and trees). The buildings’ geometry took inspiration from typical buildings in Stockholm by using reference images from buildings on Valhallavägen, using the Street View function of Google Earth¹, which would be the basis for the whole

1Google Earth. https://www.google.com/earth

(9)

virtual neighborhood.

1.1 Aim

To create a virtual neighborhood using 3D modeling and texturing that can be used to evaluate urban perception of it. The neighborhood should consist of a street with sidewalks, buildings, greenery (trees and bushes) and additional objects such as streetlights, garbage cans and benches in order to make the neighborhood feel more natural. The virtual neighborhood will make use of game technology that utilizes procedural generation methods for making real-time adjustments for the properties: building height, greenery, dirt, graffiti and window breakage.

1.2 Objective

To use the virtual neighborhood to evaluate what significance these properties have, and how they rank against each other, and consequently to find out if the results show any potential promise for using game technology for these kinds of evaluations on the perception of urban areas.

1.3 Delimitations

Since this project can be very big and becomes much more time consuming implementation wise with each added parameter, some delimitations were needed. The parameters are limited to building height (number of floors), dirt, graffiti, broken windows and greenery (trees and bushes). These parameters were picked based on research on urban perception, which will be covered more in detail in section 2.2.

Furthermore the time of day will be constant, i.e. the lighting settings will be constant. There will be no human agents or cars. Audio components will not be used, as only visual cues and head movements during the virtual reality tests will be utilized.

1.4 Methodology

Using game technology for neighborhood design and evaluation of the human perception of it is a very cost-effective way for production (building the neighborhood and buildings), and for making adjustments to an existing virtual environment (adding graffiti for example). Using procedural generation replicable results can be generated while still having the option for real-time modifications. Variations of the same neighborhood can easily be modified without using extra resources, something

(10)

that is not available in the real world. Game technology and procedural generation will therefore be the chosen methodology for this thesis project.

1.5 Structure

The thesis will start of with some background information on procedural generation, followed by earlier research and theories on urban perception, as well as going over the MIT Place Pulse²project - a project using visual surveys to measure urban perception - that the survey in this thesis is based on (MIT Place Pulse is covered in more detail in section 2.2.1).

After that follows the implementation starting with 3D modeling where buildings are modeled and divided, texturing and procedural generation for adding dirt and graffiti for example. Lastly the VR functionality and support will be handled, setting it up for VR testing.

Lastly follows the evaluation method were the forming of the survey and VR interviews will be treated, as well as the results and discussion of it.

2Place Pulse, MIT. http://pulse.media.mit.edu/

(11)

(12)

Background

Before starting with creation of the virtual environment and designing an evaluation method, some theory on the subject of procedural generation and urban perception were required, which are presented below.

2.1 Procedural Generation

Procedural generation is not a new phenomenon. As a method it has been utilized since long before video games were graphically oriented. In 1978 Beneath Apple Manor¹ was released, were procedural generation was used as a way to construct dungeons for its ASCII- and tile-based system. As the earliest graphics cards on computers were very limited due to memory constraints, and there was not enough space to store artwork and pre-made levels, the content had to be generated live while playing by using algorithms. With newer graphics cards this is not really an issue anymore, but is still used for other reasons. Procedural generation opens up the possibility of randomness in maps, levels, characters and quests that will be unique for each play through. More modern examples include the sandbox game Minecraft² and the upcoming space exploration game No Man’s Sky³, where the latter features an infinite procedurally generated galaxy where every planet is different from the next.

Procedural methods have been utilized extensively as a method for creating natural phenomena like trees and plants and in CityGen[9], theses methods are utilized to create a complex net of road networks that serves as a forming of cities and urban neighborhoods. Another motivation for using procedurally generated content is due to the problem of expanding content areas, as it typically demands more graphical artists in order to create detailed models. Using mathematical methods and algorithms that operate under a specific set of rules (i.e. rotational constraints

1Beneath Apple Manor Wikipedia Page. https://en.wikipedia.org/wiki/Beneath_Apple_Manor

2Minecraft. https://minecraft.net/en/

3No Man’s Sky Wikipedia Page. https://en.wikipedia.org/wiki/No_Man%27s_Sky

(13)

or probability of turning left on extrusions), much of the work can instead be automatically generated.

The Esri City Engine⁴is a tool used for urban planning, architecture, and design by procedurally generating content. It was developed at ETH Zurich (Swiss Federal Institute of Technology in Zurich) by Pascal Muller, as a part of his research[18].

The system constructs a roadmap using L-systems (often used for creating complex shapes in tree branches and bushes) with input from different images containing information about land-water boundaries and population density. By using this information, the road system could be subdivided realistically to mimic the actual area from the images, and pre-fabricated building models (not generated) could be placed accordingly. In his later research [12] the focus was on the building part, where once again mathematical rules - or "shape grammars" - were used to create the content. The buildings were created using an iterative process, firstly creating a very crude model and with each next step adding more and more detail. The result is a completely procedurally generated city that is also photorealistic while demanding very little in regards of input from the user.

Another example of using procedural generation to create buildings is the mod- eling tool developed in the paper Instant Architecture[23]. The tool automatically models architecture by using parametric set grammars based on shapes. It serves as a interactive urban planning application tool for urban reconstruction, as well as new construction, where the urban planners as well as potential residents, can explore different design choices by simply adjusting a set of parameters. The gram- mar that the tool uses 10 basic shapes, where the cuboid, cylinder and prism shapes have defined split rules. By utilizing these shapes complex buildings can be created - that can look different for each generation - as seen in Fig.2.1

Figure 2.1: Example buildings created in Instant Architecture[23]

As creating buildings manually can be a very time consuming and costly - especially when doing it on a larger scale - having the ability to generate them proce-

4Esri City Engine. http://www.esri.com/software/cityengine

(14)

Figure 2.2: Image of how buildings are divided into parts in Instant Architecture[23]

durally is a great advantage. The Unreal Engine⁵ has even implemented a built-in ProcBuilding system⁶- a special type of volume - into their developer kit UDK that allows the user to quickly design and edit city buildings by adjusting parameters while making use of different rulesets regarding shapes. Commencing with a simple cube, using the rulesets for example repetition, corner angles, and building levels, a base structure for a whole building can be generated. By setting up rules for specific meshes, like using a specific window for the top floor or using decor on all of the fa- cade edges, a realistic building can be created with a pretty low effort from the user.

Using procedural generation on additional properties, such as the ones being evaluated in this thesis (building height, dirt, graffiti, greenery and building height), a full neighborhood can be generated with respect to parameters controlled by user input, while simultaneously following rules regarding aspects such as amount, position, and rotation.

2.2 Urban Perception

A neighborhood can be divided into many different components that together make up the general perception of it. These components range from neighborhood layout, building aesthetics, and greenery to cleanliness, graffiti, and littering. As we receive a lot of information about an area directly from the environment, which subsequently affects our general perception and feelings towards it, urban perception is important for several fields such as urban design, regional planning and landscape architecture [2].

5Unreal Engine. https://www.unrealengine.com

6Unreal Engine: Procedural Buildings. https://udn.epicgames.com/Three/ProceduralBuildings.html

(15)

Figure 2.3: Example of a rule set regarding repetition in the ProcBuilding system⁶

Figure 2.4: The node-based system where base meshes and the rule sets can be adjusted⁶

2.2.1 Place-Pulse

Place Pulse⁷is a project by MIT⁸that aims to quantitatively measure urban perception. It is done by crowdsourcing visual surveys that are anonymously filled out by using pairwise comparisons of images, followed by evaluative questions like "Which place looks safer?" or "Which place looks more upper-class", as seen in Fig.2.5.

Starting out with images from only four cities (Boston, New York City, Salzburg and Linz), the project has grown a lot and now includes 56 cities, as well as machine learning for better pattern recognition captured by the participation data.[20]. The goal is a create a measurement tool of perceived safety of cities, and the impact of different visual elements of the streetscapes, and how they can influence urban planning and the architectural community[13].

7Place Pulse, MIT. http://pulse.media.mit.edu/

8Massachusetts Institute of Technology. http://web.mit.edu/

(16)

Figure 2.5: Example of the MIT Place Pulse⁷

2.2.2 Broken Windows Theory

The Broken Windows Theory (BWT) is a criminological theory introduced by James Q. Wilson and George L. Kelling in 1982 [22]. The theory is about the effect of disorder and vandalism on additional crime, as well as anti-social behavior in urban environments. The BWT states that maintaining a neighborhood and actively monitoring to prevent smaller crimes such as vandalism and littering helps to create an atmosphere of lawfulness and order, which in turn prevents more serious crimes from taking place. An example is given describing the theory:

Consider a building with a few broken windows. If the windows are not repaired, the tendency is for vandals to break a few more windows.

Eventually, they may even break into the building, and if it is unoccupied, perhaps become squatters or light fires inside.

A successful strategy for preventing vandalism is said to be addressing the prob- lems when they are small [8]. For instance repairing windows within a short times- pan will decrease the likelihood of people breaking more windows or doing any further damage to the environment.

Residents’ negligence of decay in their community - such as broken windows that goes unrepaired - signifies a lack of concern for the community [17]. This negligence is described as a sign that the society has accepted this disorder. However, informal social controls can be a very effective strategy to prevent unruly behavior.

Community policing measures through informal social control exercised through ev- eryday relationships and institutions is according to Garland[6] more effective than legal sanctions. When the community is left in a state of decay, the sense of a law enforcing community is lost with it.

(17)

The concept of fear is a crucial part of the broken windows theory, and according to Ranasinghe [19] even the whole foundation of the theory. As perception of disorder rise, fear is subsequently elevated as well. This in turn creates a social pattern that according to Ranasinghhe, tears the social fabric of a community apart, and makes the residents feel disconnected to their community.

When in an unfamiliar urban environment where there is nobody around, the social norms are not obvious. This results in the individual looking for signals within the environment as to the social norms in the setting and the risk if getting caught violating those norms[19]. One of those signals is the area’s general appearance.

An environment that is perceived to not be maintained sends the signal that the area is not monitored. Subsequently this results in that the criminal’s behavior has a lower risk of being detected. It is not the broken windows themselves that is the important part, but the actual message that they send out as they symbolize vulnerability and a lack of cohesiveness of the community.

The Broken Windows Theory has also been influential politically. It was cited as a justification for the "Quality-of-life Initiative" in New York in the mid 1990s, an order and maintenance strategy focusing on signs of disorder and petty crime, as graffiti was removed, streets were swept and signs of vandalism cleared. After the introduction of the initiative, petty crime rates dropped[20], and other approaches based on the BWT has been adopted worldwide. Although being a popular theory, it is however controversial as providing evidence to prove or disprove the theory has been difficult as longitudal studies have both argued in favor and against the BWT[7]. However during the field tests in The Spreading of Disorder [7], the mere presence of graffiti (added to the same place, as seen in Fig.2.6) more than doubled the number of people committing other smaller crimes such as littering and stealing.

Figure 2.6: A field test setting in The Spreading of Disorder where graffiti has been added in the right picture [7].

2.2.3 Greenery

Research has shown that urban trees provide a range of benefits to communities, such as moderating water runoff, and increasing property values [1]. Results from

(18)

studies that have looked at the effect of trees and other vegetation on crime have been mixed. Some have found that vegetation can increase the fear of crime [15]

[16], while others have found the exact opposite [10] [11]. Vegetation can both provide cover for the criminal, and act as a barrier of escape for the victim. View- obstructing trees that break the line of sight between the person and the windows of the building can also increase the fear of crime, since the surveillance aspect in the community is greatly reduced.

In a study [21] where subjects where shown 180 pictures of different parking lots, the results showed that the parking lots with the most vegetation were ranked as the least safe ones. This was also the case when conducting a similar study on university campuses [15]. In contrast to this, other studies have found that trees in residential areas instead reduces the fear of crime [3] [14]. Another study [11] that studied the effect of vegetation on crime in apartment buildings in Chicago found that vegetation was associated lower violent crime and lower property crime as well.

If trees are view obstructing, they could reduce the probability of the criminal being observed, which in turn would reduce the expected cost of the crime to the perpetrator. Similarly, the view-obstructing trees could reduce the effectiveness of police or other forms of surveillance - whether it being security cameras or neighbors - since the line of view is compromised.

In another study [5] the relationship between trees and crime was conducted using 2,813 single-family homes. The results stated that the number of trees on a lot was associated with increased crime occurrence, whereas having trees fronting a house were associated with decreased crime occurrence. This increase is explained - as in former mentioned studies - by the fact that trees could provide cover for crim- inals and thus increasing crime occurrence, and that the trees obstructs the field of view. The decrease on the other hand is explained that the trees would reduce crime is explained as trees might encourage people so spend more time in public places since it brings up the aesthetic quality of the neighborhood and makes them more desirable [4], which would increase the chance of a criminal getting spotted by the people in the vicinity.

Attributes of a neighborhood may provide information about the effectiveness of authority [22], and trees is an attribute that may indicate that the community is more cared for[5].

With the topics of procedural generation and urban perception handled, the next step is to do the actual implementation of a procedurally generated virtual neighborhood that has adjustable parameters regarding building height, levels of dirt, graffiti and broken windows, as well as amount of greenery. This will be covered in the next chapter: Chapter 3 - Implementation.

(19)

(20)

Implementation

With having treated both procedural generation and urban perception, the virtual neighborhood now had to be implemented. This was done by using modeling in 3D software, texturing in photo editing software, scripting in the Unity3D¹ game engine, and ultimately implementing VR compatibility for VR testing. The methods are described in this chapter.

3.1 3D Modelling

Since the goal was to have buildings that resemble typical Stockholm buildings, the function Street View in Google Earth was used to get a ground level perspective of Karlaplan in Stockholm. Screenshots were taken and imported as a background image in the modeling software Blender². The geometry of the buildings where created by firstly creating a plane that fitted around the main exterior of the building. The plane was subdivided and extruded to fit protruding areas. The more detailed parts - such as windows, balconies, and decor - where carefully traced as a new part of geometry. When a building was done, it was firstly divided for each floor and then for each apartment window so that different buildings, with different dimensions, could be generated using the reference building. A total of five different reference buildings - tracing 5 different Stockholm buildings - were created. One of them can be seen in fig. 3.1.

Cube projection³ was the main method of unwrapping the main part of building meshes, as it applies the textures evenly throughout all the sub parts, and can all share the same material - reducing computational costs. Since the bottom parts were later on supposed to have the option of getting assigned even more dirt (since dirt usually accumulate where the building meets the street) as well as graffiti, they

1Unity Game Engine. https://www.unity3d.com/

2Blender. http://www.blender.org

3UVs. https://www.blender.org/manual/editors/uv_image/uv_editing/unwrapping.html#cube

(21)

Figure 3.1: Process of tracing a simplistic Google Earth model, adding geometry and lastly splitting the building into smaller parts.

were each assigned their own material.

The street and sidewalk, as seen in fig. 3.2, was modeled after the same street pictures that the buildings were taken from. When the scene was put together it felt a bit hollow, and according to the pictures there were some benches, trashcans and lighting poles along the sidewalk. These things were not part of the evaluation, but since they make the scene feel a bit more natural, they were modeled and brought into the scene as well. The street itself (road and sidewalks) was fixed in regards to size, and was not affected in any way by the user input. The street lights and benches were also fixed into the sidewalks.

Figure 3.2: Part of the street used (untextured)

Since the modeling and texturing of buildings was more time-consuming than was originally planned for, free tree⁴ and bush⁵ assets were downloaded from the Unity Asset Store.

4Birch Tree Pack. Publisher: Works for Fun. https://www.assetstore.unity3d.com/en/#!/content/49093

5Yughues Free Bushes. Publisher: Nobiax/Yughues.https://www.assetstore.unity3d.com/en/#!/content/13168

(22)

3.2 Texturing

The majority of the textures were created in Adobe Photoshop⁶. For example: a concrete texture was made by using the cloud function, adding noise, adding difference clouds to the alpha channel, and then adding lighting effects. Some (free) textures - like the brick one - were downloaded from textures.com due to being more complex to produce by myself while getting a realistic result.

Since using a game engine with a scene that is required to run smoothly, and with increasing the detail of the meshes by adding more vertices and geometry subsequently increases the processing power required by the computer when running the scene, some techniques were used in order to fake the detailed geometry.

As only using a single image texture - brick for example - makes a mesh appear really flat and unnatural, normal maps can be used to fake complex geometry by faking lighting. For example, having a mesh consisting of a single plane (4 vertices, two triangles) that utilized both a brick texture and its affiliated normal map texture the plane can appear to have rather complex geometry while keeping processing power costs at a minimum as seen in the figure below.

Figure 3.3: A four vertices plane without and with normal map

Other techniques included occlusion mapping (as seen in Fig.3.4) that provide information about which areas that should receive high or low light, and specular mapping that determines the shininess of the object, as well as highlighting certain colors.

6Adobe Photoshop. http://www.adobe.com/products/photoshop.html

(23)

Figure 3.4: Occlusion map of a brick texture, and the same one with higher contrast.

The normal, occlusion and specular map were generated by using the software CrazyBump⁷, and only the image textures as input. Since the texture of windows does not really contain any geometric information in relation to bumps, the breakage of windows could be added directly as a detail mask and detail texture on top of the original window texture without any specific computation after being added procedurally. The result is a texture that has blended the original window texture with the cracked window texture.

The dirt for the buildings were created procedurally through scripting which will be described more in detail in the following section on procedural generation (3.3).

Figure 3.5: Original texture to the left. Only perlin noise texture applied to the right.

7CrazyBump. http://http://www.crazybump.com

(24)

Figure 3.6: Two different amounts of dirt applied using both the perlin noise texture and the high contrasted occlusion map

Using the occlusion maps to control the areas that the dirt would be applied to, the dirt was added more naturally as where the most dirt most likely would build up, instead of just applying the same amount over the whole material which would more resemble spray-paint (see Fig. 3.5) than naturally accumulated dirt (see Fig.

3.6).

The simpler un-colored graffiti images were created by writing random phrases in Photoshop using free graffiti fonts. A total of 8 were created and imported into the game engine. The more detailed graffiti textures were created using a free online graffiti generator⁸, and there was a total of 8 different ones later on brought into Unity. Examples of simple and more complex graffiti can be seen in Fig. 3.7.

Figure 3.7: Example of a simpler and more complex graffiti texture that were used.

8Graffiti Creator. http://www.graffiticreator.net

(25)

3.3 Procedural Generation

In order to make the dirt application more varied, a script that could produce a black and white perlin noise texture was created. The bright parts of the perlin noise texture are then removed, making those pixels completely transparent. This way each generation of the perlin noise will result in a different dirt texture for the main walls of the building. The texture is then applied as a detail mask/texture in the standard shader of Unity. The amount of dirt that would be applied to the main fa- cade (floor 2 and up) used the occlusion map method of applying the dirt, as seen in Fig.3.6, where the input from the user would decide the opacity the dirt would have.

In order to be able to differentiate the height of the buildings - more precisely the number of floors - each building had to be broken down into several smaller parts, as described in the 3D modeling section. A script was then created that accepts user input of between four and eight stories tall, and then generating it, as seen in figure 3.8. The pieces that were affected by the script was every floor except for the bottom and top floors, since they are essential in making a building actually look natural.

Figure 3.8: Three different generations of buildings

All windows were tagged and put into a list. The user input that ranged from 0-100% would then decide the amount of windows that would get a breakage texture applied to them. The amount was final, but the specific windows that would receive the breakage texture were randomized. The same procedure was done for the low parts of the buildings that would be able to receive extra dirt and graffiti.

Potential positions for both bushes and trees were manually placed within the neighborhood scene. The potential tree positions that would then receive a ran- domly chosen tree was then calculated simply by multiplying the percentage from the slider input by all possible tree positions. Bush placement was used in the same way.

After adding all the parts together - modeling, texturing, and scripting - the

(26)

virtual neighborhood could then be generated procedurally based on chosen parameter values. To be able to control the parameters more effectively, a user interface was implemented with simple slider controls, as seen in fig. 3.9. The user simply inputs values using the mouse or keyboard, then presses generate and all the values will be passed on to the different scripts. The only part left in the implementation was adding the option to be able to use the virtual reality headset Oculus Rift, as described in the next section.

Figure 3.9: Image of the user interface

3.4 Virtual Reality

Oculus provides a software development kit (SDK)⁹ on their developer website, that is easily added to an existing project by just downloading the package and then im- porting in Unity into an existing scene, replacing the standard camera.

Since I do not own an Oculus Rift myself, but have access to an Android phone, the testing was first done by using the Google VR SDK¹⁰, a virtual reality SDK that works with an Android phone and a Google Cardboard (simplistic VR headset assembled by cardboard and two lenses that you slide your Android phone into).

This was used when scripting the walkthrough in order to make it feel natural in the sense of view height, walking speed and rotation.

9Oculus SDK. https://developer.oculus.com/downloads/

10Google VR SDK. https://developers.google.com/vr/

(27)

Having a working implementation of the virtual neighborhood, that through the user interface can vary the intensity of the parameters broken windows, graffiti, dirt, greenery and building height, an evaluation method for testing it was needed. An image based comparison survey inspired by the MIT Place Pulse (covered in chapter 2.2.1), as well as a virtual reality walkthrough - in order to get more detailed responses from the participants - was utilized which will be covered in the next chapter.

(28)

Evaluation

Having the procedurally generated virtual neighborhood done, a scripted walkthrough was used in a VR environment in order to evaluate it. Additionally a survey inspired by the MIT Place Pulse was used for more quantitative results.

This chapter begins with the implementation of these two methods, followed by their results, and finally ending with analysis/discussion over the results gained from the survey and VR interviews.

4.1 Evaluation Methodology

4.1.1 Survey Implementation

A survey with in-game screenshots was created using the same method as utilized in the MIT Place Pulse project, as mentioned in section 2.2.1. This method opens up the possibility of pinpointing the perceived difference between specific parameters, as well as gaining quantitative data that would serve this thesis problem statement well. Google Forms was used and each question consisted of two pictures where the participant were to chose the picture that conveyed the feeling of being the most safe area - or if they conveyed the same feeling. To get the screenshots multiple scenes had to be generated in the Unity game engine. Multiple scenes were generated with different settings for amount of dirt, graffiti, broken windows and greenery while in-game using the sliders controlling the parameters. Since the scenes are generated by using procedural/modular generation with a random number generator to figure out where to instantiate trees for example, two scenes could differ somewhat even though using the same input on the different parameters. Therefore a lot of scenes had to be generated in order to get the two scenes to be very similar, with the only difference being the one part that was being investigated. For example, to get two scenes with moderate amounts of dirt, graffiti and dirt (that would also look really similar) but with different amounts of trees, a fair amount of scene generations were needed. In-game screenshots were then taken, and brought into the survey. The form consisted of 19 questions regarding the user’s perceived safety between two pictures, where the participant would choose whether if one of the pictures was

(29)

regarded to feel like a safer environment. A safer environment was defined as the environment that would make you feel more secure and to be less associated with crime. The survey can be viewed in its entirety in Appendix A.

The survey was distributed on the social news-networking site Reddit¹, as well as through social networking site Facebook². The survey got a total of 153 responses.

Table 4.1: Table of intensity settings for the different properties that were used in the survey and walkthrough. "L" denotes low intensity, "M" denotes medium intensity, and "H" denotes high intensity. The dash "-" denotes an absence, e.g.

Greenery(-) means that there is no greenery at all in the scene, and Greenery(H) means a maximum amount of greenery.

Properties Intensity Settings

Dirt -/M

Graffiti -/H

Broken Windows -/H

Greenery -/L/M/H

Building Height L/M/H

Examples of different settings can be seen below in fig. 4.1 and fig. 4.2

Figure 4.1: Example with building height(L), dirt(-), graffiti(-), greenery(-) and broken windows(-).

4.1.2 VR Test and Interviews Implementation

Since the survey only consists of choosing between three choices per question, and does not open up for longer and more detailed answers, interviews were also used. A test was constructed where the participant would enter the neighborhood virtually,

1Reddit. https://www.reddit.com/

2Facebook. https://www.facebook.com/

(30)

Figure 4.2: Example with building height(H), dirt(H), graffiti(H), broken windows(H), greenery(H)

using VR since it tends to increase the immersion for the player, as opposed to using a regular computer or TV-screen. The method used was semi-structured, since it allows follow-up questions based on what the participant says in a particular moment.

In order for the test to really focus on the visuals cues, the movement was scripted. Using checkpoints along the sidewalk, a 10 second long walkthrough was created, where users could not move by themselves, but was rather guided by the script. The user could however control the rotation of the in-game camera by turning and moving his/her head and still be able to focus on details.

Participants were recruited using a mail list of students at KTH taking a computer graphics class that my supervisor Christopher Peters is teaching. It resulted in 4 participants and the VR test consisted of 5 walkthroughs with the different settings as written in table 4.2

Table 4.2: Properties of the scenarios in the VR tests, where "L" and "H" refers to low respectively high intensity, or height in the building height column.

Scenario Label Building Height Dirt Graffiti Broken Windows Greenery

SC1 H - - - -

SC2 H H - - L

SC3 H H H H L

SC4 H H H H H

SC5 L H H H H

(31)

4.2 Survey Results

Dirt, graffiti, and broken windows were used with high intensity or very low intensity. Greenery was used in four settings. No greenery at all, low intensity (denoted

"L") where there was very few trees and bushes and none of them view obscuring, medium intensity (denoted "M") where there were some trees where some of them could be view obscuring, and high intensity where all possible tree and bush positions in the scene were utilized where the view was very obscured. Building height was used with low, medium or high intensity (meaning from 4 to 8 floors). Set- tings for height were mainly used with a clean scene as in the first three questions, or in the last questions with a more intense scene with dirt, graffiti and a lot of trees.

Figure 4.3: Results for question 1 to 3. Q1 refers to question 1 in the survey as seen in Appendix A, where "Left" and "Right" refer to the pairwise comparison of the two images. "Same" refers to the option that the participants did not feel any differences between the two images. In Q1 for example, the left picture had tall buildings, and the right picture had low buildings. "L", "M", and "H" refers to intensity settings of low, medium, and high intensity. The dash "-" refers to non-existent.

The results of the survey and the first three questions focusing on building height showed that there was no clear favoring of a particular building height, in regards to conveying the feeling of safety. During the first three questions where height was the only difference between the two pictures, the low height setting was slightly favored over the tall (+5.2 %), and the medium height was favored over both the tall and low heights (+5.9 %, respectively (+2.6 %). Between 41.2 and 52.3 percent did not feel that the height made any difference at all.

The following three questions which dealt with no trees, some trees, or a lot of trees that covered the other side of the road, there was a great favor of choosing the ones with the most trees. 33.3 % more participants chose the one with some trees over the one with no trees. 61.4 % more participants chose the one with a lot of trees over the one with few trees. 58.1 % more participants chose the one with a lot of trees over the one with no trees. Only 12.4 - 13.1 % did not think that there was any difference between a lot of trees and some trees, and a lot of trees and no trees. Between no trees at all and few trees the indifferent choice went up to 35.3 %.

(32)

Figure 4.4: Results for question 4 to 6. The building height was reset to medium height for all the remaining questions until written otherwise.

Figure 4.5: Results for question 7 to 9.

Question 7 regarding dirty buildings or clean buildings, 22.2 % felt that the dirtier building felt safer and 54.9 % felt that the cleaner one did. 22.9 % did not think that the dirt made any difference.

Question 8, where the comparison was between broken windows and no broken windows, showed the so far greatest difference in choice. A great majority of 71.9 % picked the one with no broken windows, and only 1.9 % picked the one with broken windows on it. 26.2 % did not think it made any difference.

Question 9 regarding buildings with graffiti or no graffiti, 77.1 % picked the one without graffiti and 6.5 % picked the one with graffiti. 16.4 % did not think that the graffiti made any difference.

Question 10 was between a dirty building and a clean building with graffiti.

There was a strong favor in dirt being the one that conveyed the most unsafe area with 64.7 %, whereas the graffiti one got 20.9 %. 14.4 % picked the option that there was no favor in any of them.

Question 11 was between a neighborhood with broken windows, and one with graffiti. A slightly bigger portion picked the option that the graffiti version conveyed the most unsafe feeling with 45.1 %, than the one with broken windows at 38.5 %.

16.4 % did not favor any one of them.

(33)

Figure 4.6: Results from question 10 to 12

Question 12 was between buildings with broken windows and dirty buildings.

More than twice as many picked the option that the version with dirty buildings conveyed the most unsafe feeling at 58.2 %, whereas 23.5 % picked the one with broken windows. 18.3 % did not have favor any of them.

Figure 4.7: Results for questions 13 to 15

Question 13 was between a clean neighborhood with no greenery but with broken windows, and a neighborhood that had dirty buildings, broken windows and a lot of greenery. There was roughly twice as many that picked the latter at 64.1 %, whereas the first got 26.8 %. 9.1 % did not have any favor.

Question 14 was between a neighborhood that had dirty buildings with graffiti and broken windows with a lot of greenery, and a neighborhood with dirty buildings and graffiti but without broken windows and with less greenery. The results were roughly the same between the two scenes, where the first one got 35.9 % and the second 30.7 %. 33.4 % picked the option that they conveyed the same feeling.

Question 15 was between a neighborhood with dirty buildings, broken windows and a lot of greenery, and a neighborhood with dirty buildings, graffiti and a lot of greenery. 49.7 % thought that the first one communicated a safer neighborhood, and 32.7 % thought that the second one did. 17.6 % thought that they were about the same.

Question 16 was between a neighborhood with dirty buildings with graffiti and

(34)

broken windows, and no greenery, and the same neighborhood but with a lot of greenery. There was a big favor for the one with greenery at 58.8 %, versus the one with no greenery which got 13.1 %. 28.1 % did not favor any of them.

Question 17 was between two neighborhoods with the same amount of dirt, graffiti and broken windows. The first had a fair amount of greenery, and the second one had rich greenery. The second one was in great favor at 56.2 %, whereas the first one got 14.4 %. 29.4 % did not favor any of them.

Question 18 had the same setup as question 17, except the first one has some greenery, and the second one no greenery at all. 49 % picked the option that the one with some greenery felt safer, and 16.3 % that the one with no greenery did.

34.7 % did not favor any of them.

Question 19 had two neighborhoods with dirty buildings, graffiti, broken windows and a fair amount of greenery. The first one however had tall buildings, and the second one had very low buildings. 32.7 % thought that the one with lower buildings felt safer, and 22.9 % thought that the one with taller buildings did. Most participants did not favor any of them, as this choice got 44.4 % of the votes.

(35)

4.3 Virtual Reality Interviews

The four participants will hereafter be referred to as P1 to P4, and the 5 different scenarios/walkthroughs will be called SC1-SC5 as described in table 4.2

Regarding the difference between SC1 and SC2 it was said that:

It felt more immersive, but I was not taken aback by the fact that there was some dirt on the walls. - P1

I did not feel any difference between the two scenes at all. - P2

I did not feel any big difference, perhaps the buildings are a bit dirtier.

Some trees as well, but the dirt was the first thing I noticed. But it did not make that much difference. - P3

Well, I notice some dirt on the walls. And the trees were not there before either. The trees, I think, make the neighborhood feel a bit more pleasant. I would of course prefer if there were no dirt on the buildings, but it does not make that much of a difference for me. - P4

Regarding the difference between SC2 and SC3 the following was said:

Now I felt a big difference. It feels a lot more like a ghetto. When looking at it I think: "this is not a nice neighborhood". At a long distance it is pretty difficult to notice the detailed differences since it is quite pixelated.

But when looking a bit closer the overall impressions is that it is probably not a neighborhood that no one really cares about, since it is both graffiti and broken windows. If there were only graffiti - and no broken windows - I would probably think, "this is something that they will be removing soon". - P1

I noticed a lot of graffiti, and that would certainly lower the housing prices. The fact that there is so much graffiti left up, I think indicates that it is not so well-managed or well supervised. Now I noticed some broken windows as well. It feels a lot more unsettling now, definitely. - P2

The graffiti was the first thing I noticed, but also that the windows are broken. It feels more like a slum, definitely, despite the fact that it is the same neighborhood. The graffiti is more noticeable, or at least I noticed that first. But I think the windows have a greater impact. - P3

The graffiti and the broken windows really drags down the overall im- pression I think. The graffiti is more noticeable on a longer distance.

But the broken windows make the neighborhood feel almost abandoned.

Or that no one cares about it. - P4

(36)

I did not really notice any difference at first. My impression was pretty much the same as with the last scene. - P1

I do not know if this changes anything for me. There are more trees, but the damages are still the same level. And I think it is still as worrisome, even though there are more trees. - P2

I do not really notice any difference between this and the last scene. - P3

There are more trees now and it makes it feel a bit more crowded. That it is a bit more difficult to see to the other side of the street. I think I would prefer if there were less of them, since the sight of the neighborhood is pretty low. - P4

It feels like the graffiti and everything does not fit together with the scene anymore. It feels absurd with this much breakage, since the neighborhood now feels more like a small town area, than an urban area. It does not feel real anymore, because I have never been to an area that looks like this before. When there were taller buildings it felt like the graffiti and broken windows fitted together more. - P1

I can not really tell what the difference is between this and the last scene.

[After telling P2 that the buildings are a lot lower] -> Oh, yes, that is correct. I did not notice that. But I do not think that makes any difference for me. It is still the broken windows and the graffiti that has significance. I do not know what is worse, the graffiti or the broken windows, but the combination is very alarming. - P2

The houses are lower, right? It feels a lot brighter now. The tall build- ings gave more of a confined feeling, since they almost felt like walls in a room. With the lower buildings it feels a lot brighter and more open.

- P3

It feels like the area is more open, and less claustrophobic than before.

- P4

(37)

4.4 Analysis and Discussion

The results of the survey clearly show that trees are preferable when conveying a sense of safety. What was somewhat surprising is that it did not seem to matter at all that the trees were view obscuring, making it practically impossible to see to the other side of the street or the first floor windows. Instead it seemed that the more trees that were present, the better. It is possible that this is due to trees being seen as an attribute that indicate that the community is cared for [5] and therefore has a higher sense of authority as mentioned in the background. Or that the trees brings up the aesthetic qualities of the neighborhood and in turn would make the neighborhood more desirable and possibly encourage people to spend more time in the area[4]. Most participants of the VR test did not feel that the trees by themselves made any particular difference at all, and were more bothered by the dirt, graffiti and broken windows. Only one participant mentioned that it felt a bit more crowded and was bothered by the line of sight being compromised. The indifference when choosing between pictures that had either no trees, medium or high amount of trees, was very low at only 12.4-13.1% and thus signifying that trees and greenery are highly relevant factors.

When a different level of dirt was the only significant factor between the two pictures, the majority picked the cleanest one. This was to be expected since having dirt on buildings can signify that the neighborhood is poorly looked after, and as a result communicates a feeling that it also supervised and/or has a poor sense of community within the residents [17], which might explain that the dirt added to increasing the feeling of the area feeling more unsafe. A low amount (22.9%) - but higher than with the tree questions - did not feel that the difference in amount of dirt made any difference at all. This might be due to dirt usually being perceived as more of a temporary issue that will probably be dealt with in a relatively short amount of time (or just fixed through rainy weather). What was interesting however is that the significance of dirt was substantially lower in the VR tests as most participants seemed to feel that the added dirt to the buildings made very little to no difference at all. During the VR tests the participants got a much closer look of the buildings, and one might think that the effect might be more palpable and therefore signaling a sense of disorder compared to the scene with clean settings.

What might explain this is that the participants of the VR test had to wait for a while between the scenes, since they had to be re-generated while the participants of the survey got a view of the two scenes at the same time, making it easier to make a comparison between them.

The broken windows made a great difference both in the survey and in the VR test. In the survey 71.9% picked the scene with broken windows as the scene conveying the feeling of being the most unsafe, and only 1.9% picking the other one.

What was interesting during the VR tests is that the broken windows was noticed rather late by most participants - where example graffiti and dirt were noticed long

(38)

before - but ultimately chosen to be the biggest concern in regards of disorder of the community. The broken windows conveyed a feeling that the community was not only poorly maintained but also that this particular damage had been present for a very long time as opposed to dirt which was more seen as a temporary problem and therefore not seen as such a big factor. A fair amount (26.2%) of survey participants chose the option that the broken windows did not make any difference. Since there was such a great difference between choosing image 1 and image 2, and the fact that it posed such a big issue for the VR test participants, chances are that the images might have been to small for the indifferent participants to actually notice that the windows were broken.

A great majority (77.1%) in the survey picked the image with graffiti on the walls, and only a low amount (16.4%) of participants did not think that it made any difference, suggesting that graffiti is a very relevant factor. When compared to dirt there was a clear majority choosing the one with graffiti, which might be - just as with the broken windows - due to being seen as more of a longtime problem.

Building height did not pose as an important factor in neither the survey nor during the VR tests. The survey results showed that the favoring always were roughly the same between two different heights images, and the indifference option always got the majority of the votes, which also suggests that building height is not a very relevant factor. The interviews showed that the building height did not make any particular difference at all, only that it made the scene a bit brighter.

What was very unexpected is that in the surveys, dirt conveyed a more unsafe feeling than both graffiti and broken windows when comparing dirt and graffiti, and dirt and broken windows. This is the exact opposite from what was found out during the interviews. During the interviews dirt and graffiti was noticed long before broken windows, which might indicate that the survey participants might have picked the one that were more prominent of the two, or that they did not even notice that the windows were broken. The dirt has a higher contrast towards the wall than the one that had graffiti, which might explain the choice of dirt over graffiti, especially if the participant’s view of the form was not maximized.

Whether having two pictures side by side for comparison being the best choice, considering that the images obviously get a smaller portion of the screen, is debat- able. This was a conscious choice however, as it is usually easier to tell a difference between two things that are in the field of view than having to scroll back and forth between two full screen pictures. Since using two images side by side for comparison, a lot of the participants screen area also went unused. Due to this, the amount of detail noticeable by a participant is obviously a lot lower than it would have been if each image would have been presented in full screen size. Also the form does not require any specific screen size, so it is possible that some of the participant could have been filling the survey out on mobile, which if not zooming in makes those results borderline useless. It should be said that just because a certain variable -

(39)

whether it being broken windows, dirt, graffiti or amount of greenery - was or was not significant in the results, since using the sample size ideally would have to be a lot bigger, one should be very careful to use the results presented in this study as generalizable. But it might at least serve as some kind of a guideline when creating a virtual street scene, or for further experiments in the area.

(40)

Conclusion

Greenery, broken windows, dirt and graffiti are all significant when it comes to the user’s perception of a virtual street scene. The properties order of significance however differed between the surveys and the interviews. The survey showed that dirt was the most significant factor when it came to conveying the feeling of the neighborhood feeling less safe, followed by graffiti and then broken windows. The interviews on the other hand showed the complete opposite, where broken windows being the most significant factor, followed by graffiti and then dirt. The interviewees picked the property that seemed to have been there the longest, which agrees with a lot of the theory presented in the background. The VR walkthroughs took approximately 20 seconds excluding time when the walkthrough had come to an end where the participants mainly looked around, so they had more time to examine the area before making a decision. The time for choosing in the survey was very likely a lot shorter, and it is possible that the participants just picked the one that showed the most contrast to the building wall. One thing that speaks for this is that most interviewees noticed the dirt before broken windows or graffiti.

The survey showed consistently that the more trees and greenery that are present, the safer the area feels. The majority of interviewees did not think the trees made any noticeable difference though. The only property that did not seem to have an impact on the user’s perception of the neighborhood was building height, as it was mainly an irrelevant factor in both surveys and interviews.

The survey and interviews showed that properties had an impact on the participants feeling of safety - with building height being the exception - which might suggest that using a virtual neighborhood to measure urban perception is a viable - and highly adjustable and cost-effective - method for evaluating urban perception.

(41)

5.1 Future Work

The virtual neighborhood presented in this thesis could be expanded greatly for further research. In this thesis, no people or cars were used, the sidewalk and main road was not affected by the dirt property, and the lighting was constant (perceived time of day). If the lighting was instead dynamic, making the streetlights (positions and amount) into an adjustable parameter could also be an interesting parameter to evaluate. Adding a garbage/litter parameter could also be implemented as it, as suggested in the Broken Windows Theory, very likely would increase the risk of more petty crime. Finally adding an option to be able to adjust street width could be an interesting parameter to examine, as having a very narrow road/sidewalk likely make the neighborhood feel more claustrophobic, and conversely having a very wide road probably would make the neighborhood feel more open and perspicuous.

Using Google Forms for the survey part of the evaluation might not have been ideal either, as there was no way to guarantee that the images were presented in full screen, even though it was encouraged in the beginning of the survey. In this thesis fear was in focus in the evaluation, but just like in the MIT Place Pulse project many other questions could be evaluated such as which place looks livelier, wealthier, more beautiful, depressing or boring.

(42)

Acknowledgments

I would like to thank all participants of the survey and VR tests, as well as Christo- pher Peters at the Computational Science and Technology (CST) and School of Computer Science and Communication (CSC) at KTH for supervision and guid- ance through this thesis.

(43)

(44)

[1] Linda M Anderson and H Ken Cordell. Residential property values improved by landscaping with trees. Southern Journal of Applied Forestry, 9(3):162–166, 1985.

[2] Donald Appleyard. Notes on urban perception and knowledge. na, 1973.

[3] Sidney Brower, Kathleen Dockett, and Ralph B Taylor. Residents perceptions of territorial features and perceived local threat. Environment and behavior, 15(4):419–437, 1983.

[4] Lawrence E Cohen and Marcus Felson. Social change and crime rate trends: A routine activity approach. American sociological review, pages 588–608, 1979.

[5] Geoffrey H Donovan and Jeffrey P Prestemon. The effect of trees on crime in portland, oregon. Environment and Behavior, 44(1):3–30, 2012.

[6] D. Garland. The Culture of Control: Crime and Social Order in Contempo- rary Society. Annali della Facoltà di Giurisprudenza di Genova. Collana di monografie. OUP Oxford, 2001.

[7] Kees Keizer, Siegwart Lindenberg, and Linda Steg. The spreading of disorder.

Science, 322(5908):1681–1685, 2008.

[8] G.L. Kelling and C.M. Coles. Fixing Broken Windows: Restoring Order And Reducing Crime In Our Communities. A Touchstone book. Free Press, 1997.

[9] George Kelly and Hugh McCabe. Citygen: An interactive system for proce- dural city generation. In Fifth International Conference on Game Design and Technology, pages 8–16, 2007.

[10] Frances E Kuo, Magdalena Bacaicoa, and William C Sullivan. Transforming inner-city landscapes trees, sense of safety, and preference. Environment and behavior, 30(1):28–59, 1998.

[11] Frances E Kuo and William C Sullivan. Environment and crime in the inner city does vegetation reduce crime? Environment and behavior, 33(3):343–367, 2001.

(45)

[12] Pascal Müller, Peter Wonka, Simon Haegler, Andreas Ulmer, and Luc Van Gool. Procedural modeling of buildings. Acm Transactions On Graphics (Tog), 25(3):614–623, 2006.

[13] Nikhil Naik, Jade Philipoom, Ramesh Raskar, and César Hidalgo. Streetscore–

predicting the perceived safety of one million streetscapes. In 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 793–799. IEEE, 2014.

[14] Jack L Nasar. A model relating visual attributes in the residential environment to fear of crime. Journal of Environmental Systems, 11(3):247–255, 1982.

[15] Jack L Nasar and Bonnie Fisher. Hot spots of fear and crime: A multi-method investigation. Journal of environmental psychology, 13(3):187–206, 1993.

[16] Jack L Nasar, Bonnie Fisher, and Margaret Grannis. Proximate physical cues to fear of crime. Landscape and urban planning, 26(1):161–178, 1993.

[17] O. Newman. Defensible Space: Crime Prevention Through Urban Design. Ar- chitecture/Urban affairs. Collier Books, 1973.

[18] Yoav IH Parish and Pascal Müller. Procedural modeling of cities. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 301–308. ACM, 2001.

[19] Prashan Ranasinghe. Jane jacobs framing of public disorder and its relation to the broken windows theory. Theoretical Criminology, page 1362480611406947, 2011.

[20] Philip Salesses, Katja Schechtner, and César A Hidalgo. The collaborative image of the city: mapping the inequality of urban perception. PloS one, 8(7):e68400, 2013.

[21] Garnett S Shaffer and LM Anderson. Perceptions of the security and attractive- ness of urban parking lots. Journal of Environmental Psychology, 5(4):311–323, 1985.

[22] G Wilson, J. Kelling. Broken Windows the police and neighborhood safety, 1982.

[23] Peter Wonka, Michael Wimmer, François Sillion, and William Ribarsky. In- stant architecture, volume 22. ACM, 2003.

(46)

Survey

(47)

dirt and greenery) in a Virtual Environment

1. Mark only one oval.

Left Right Same

1. Which place seems safer? (more secure environment for you and less likely to be associated with crime). Zoom in by holding Ctrl/CMD and press the + button if images appear small.

2. Which place seems safer?

(48)

Left Right Same

4. Which place seems safer?

5. Which place seems safer?

(49)

Left Right Same

Page 2/3

Left Picture Right Picture Same

6. Which place seems safer?

7. Which place seems safer?

(50)

9. Which place seems safer?

10. Which place seems safer?

(51)

Page 3/3

11. Which place seems safer?

12. Which place seems safer?

(52)

14. Which place seems safer?

15. Which place seems safer?

(53)

16. Which place seems safer?

17. Which place seems safer?

18. Which place seems safer?

(54)

Powered by

19. Which place seems safer?

(55)