INOM
EXAMENSARBETE TEKNIK, GRUNDNIVÅ, 15 HP
STOCKHOLM SVERIGE 2018,
Creating and Evaluating an
Interactive Visualization Tool For Crowd Trajectory Data
CHRISTINA SONEBO JOEL EKELÖF
KTH
SKOLAN FÖR ELEKTROTEKNIK OCH DATAVETENSKAP
Creating and Evaluating an Interactive Visualization Tool For Crowd Trajectory Data
CHRISTINA SONEBO JOEL EKELÖF
Date: June 6, 2018
Supervisor: Christopher Peters Examiner: Örjan Ekeberg
School of Electrical Engineering and Computer Science Swedish title: Att bygga och utvärdera ett interaktivt visualiseringsverktyg för gångbanor hos folksamlingar
Abstract
There is currently no set standard for evaluating visualization environments.
Even though the number of visualizations has increased, there is a tendency to overlook the evaluation of their usability. This thesis investigates how a visualization tool for crowd trajectory data can be made using the visualization technique of animated maps and the JavaScript library D3.js. Furthermore it explores how such a visualization tool can be evaluated according to a suggested framework for spatio-temporal data.
The developed tool uses data taken from the UCY Graphics Lab, consisting of 415 trajectories collected from a video recorded at a campus area. User evaluation was performed through a user test with a total of six participants, measuring effectiveness as completed tasks, and satisfaction as ease of use for three different amounts of trajectories. Qualitative data was recorded through using the think aloud protocol to gather feedback to further improve the im- plementation. The evaluation shows that the visualization tool is usable and effective, and that the technique of animated maps in combination with a heatmap can aid users when exploring and formulating ideas about data of this kind. It is also concluded that the framework is a possible tool to utilize when validating visualization systems for crowd trajectory data.
Keywords: Visualization techniques, D3.js, spatio-temporal data, user evalua- tion, evaluation framework, animated maps
Sammanfattning
Det finns i dagsläget ingen etablerad standard för att utvärdera visualiseringssy- stem. Även om antalet visualiseringar har ökat finns det en tendens att förbise utvärderandet av deras användbarhet. I det här arbetet undersöker vi hur ett visualiseringsverktyg för data av gångbanor hos folksamlingar kan skapas, med hjälp utav visualiseringsmetoden animated maps och JavaScript-biblioteket D3.js. Vidare undersöker vi hur det är möjligt att evaluera ett visualiserings- verktyg utefter ett givet ramverk.
Visualiseringsverktyget använder data från UCY Graphics Lab. Datan består av 415 gångbanor som är insamlade från en videoinspelning av ett campusområde.
En utvärdering genomfördes sedan med sex deltagare, där visualiseringens effektivitet och användarvänlighet mättes. Frågorna ställdes för tre olika mäng- der av gångbanor. Kvalitativa data dokumenterades genom en så kallad ”think aloud”, för att ge återkoppling och förslag på möjliga förbättringar av visu- aliseringen. Evalueringen visar på att animated maps i kombination med en heatmap kan hjälpa användare att utforska data av gångbanor hos folksamling- ar, samt att verktyget är effektivt och användbart. Det är också visat att det ramverk som användes vid evalueringen är ett möjligt verktyg för att validera visualiseringsverktyg av den typ som gjorts i det här projektet.
Contents
1 Introduction 1
1.1 Purpose . . . 1
1.2 Research Questions . . . 2
1.3 Scope . . . 2
1.4 Thesis Outline . . . 3
2 Background 4 2.1 Trajectory Data . . . 4
2.2 Crowds . . . 5
2.2.1 The Dataset . . . 5
2.3 Data Visualization . . . 6
2.3.1 The Visual Information Seeking Mantra . . . 7
2.3.2 Techniques to Visualize Trajectory Data . . . 7
2.4 User Evaluation . . . 9
2.4.1 Usability Evaluation and Common Metrics . . . 9
2.4.2 Think Aloud Protocol . . . 10
CONTENTS
2.4.3 Analytical Tasks for Trajectory Data . . . 10
2.5 Related Work . . . 11
2.5.1 Spatio-Temporal Visualizer . . . 11
2.5.2 Comparing 2D Static Maps and STC . . . 12
2.6 Tools . . . 13
2.6.1 DOM . . . 13
2.6.2 Web-development tools . . . 13
3 Method 14 3.1 The Implementation . . . 14
3.1.1 Data Cleaning and Formatting . . . 14
3.1.2 The Visualization . . . 15
3.2 Mapping Framework to User Tasks . . . 17
3.3 Pilot . . . 19
3.4 User Study . . . 19
4 Results 20 4.1 Implementation . . . 20
4.2 User Test . . . 22
5 Discussion 29 5.1 Future Work . . . 31
6 Conclusion 32
CONTENTS
Bibliography 33
A Results in Tables 36
CONTENTS
Chapter 1 Introduction
1.1 Purpose
There are several research areas where there is a need to understand and explore trajectory data of crowds. One such example is crowd simulations, where an interest lies in understanding what realistic-looking trajectories look like, for use in the special effects of movies, games and architectural walkthroughs.
There are also applications where one is less concerned with simulated virtual pedestrians but more so with the behaviour of real humans, like evacuation scenarios and surveillance footage. But generally datasets of trajectories are large and complex, and can run over extended periods of time [12]. This of course poses a challenge - how can users meant to inspect data of this nature be aided in this task?
The field of visualization research might provide an answer, as it consists of building new tools and techniques meant to improve the users cognitive process [21], and could be viewed as assisting humans with data analysis, through either creating completely new ways of visualizing data or by increasing one’s ability to interact with it.
However, the usability of such a visualization also needs to be validated. There are several different approaches to evaluate the usability of visualization en- vironments, and there is currently no established standard. Even though the number of visualizations has increased, there is a tendency to overlook the eval-
1
2 CHAPTER 1. INTRODUCTION
uation of their usability [8]. There are studies where the technique "validation through awesome example" [10] still is applied - convincing the reader of its viability by simply presenting pictures of the visualization and its features.
Thus the purpose with this thesis is twofold - first an attempt is made to build a visualization tool which can aid users in their exploration of crowd trajectory data, but also evaluate it according to a framework.
1.2 Research Questions
The following research questions will be investigated in this thesis:
• How can D3.js be used to create a visualization of crowd trajectories?
• How can such a visualization be evaluated?
1.3 Scope
The implementation built is based on techniques from cartography and ge- ographical information systems. However it does not utilize GPS data and focuses on the movement of crowds. The dataset used in this work is small and originates from a video file. The video captures the movement of people walking at a campus area and the data consists of 415 trajectories in total. This is also the maximum amount of trajectories that the tool will be tested for.
As for the validation of the final product a user test will be conducted using a framework proposed by Adrienko et al. [16] for analytical tasks of spatio- temporal data. This framework contains four different categories of tasks, whereof three of these categories are to be tested in this work. The metrics measured will be effectiveness as completed tasks and satisfaction as ease of use, where users rate the experienced difficulty of the tasks on a scale from 1 to 5, where 1 is easy and 5 is difficult.
CHAPTER 1. INTRODUCTION 3
1.4 Thesis Outline
Chapter 2 describes relevant research and techniques used to create visualiza- tions of trajectories. It also presents the dataset and tools used in this project.
Chapter 3 describes the methods used when creating the visualization tool and how it was evaluated.
Chapter 4 presents the results generated by the project and the user study.
In chapter 5 the results are discussed with regards to the research questions.
Future work and possible improvements are also provided.
Chapter 6 presents the conclusions that were drawn from this project.
Lastly, there is a bibliography of sources and an appendix with the results in tables.
Chapter 2 Background
In this section relevant background information is introduced. Firstly trajectory data and crowds are defined. Then topics related to the data visualization are presented, such as visualization techniques and design guidelines together with theory and techniques related to user evaluation. Lastly, related works are presented.
2.1 Trajectory Data
Trajectory data is spatio-temporal - belonging to both space and time. A trajectory can be viewed as the trace of a moving object, a path through space as a function of time. Examples of moving objects could be anything from people to particles - their commonality being that they are entities with positioning or geometrical properties that change over time [19]. The data of trajectories is typically represented by a set of chronologically ordered location points, P = hxn, yn, tni where xi, yiare geographical coordinates at time ti, and n is the total number of elements in the series [15]. When put together, these create a trajectory T = {P1, P2, ..., Pn}.
It is also possible that the data contains information besides the location points themselves. These attributes are either derived from or associated with the data and is often referred to as thematic [12]. Examples of such attributes could be category, the speed of an object at a given time, or what direction the object is
4
CHAPTER 2. BACKGROUND 5
facing.
2.2 Crowds
There are several definitions attempting to describe what a crowd is. However, Challenger et al.[7] mention that there are several common characteristics between these crowd definitions - stating that a crowd is a large group of people that are gathered at a specific location at a specific time and that they display common goals and behavior.
They also list a set of key criteria which may characterize a crowd:
• Size - there should be a sizeable gathering of people.
• Density - crowd members should be co-located in a particular area, with a sufficient density distribution.
• Time - individuals should typically come together in a specific location for a specific purpose over a measurable amount of time.
• Collectivity - crowd members should share a social identity, common goals and interests, and act in a coherent manner.
• Novelty - individuals should be able to act in a coherent manner, despite coming together in an unfamiliar situation.
2.2.1 The Dataset
The chosen dataset for this thesis has its origins from the UCY Computer Graphics Lab [22]. The material consists of a video of people walking through a campus area, as seen in Figure 2.1. This video file is accompanied by a .vsp file containing a series of coordinates, paired with the frame in which they where captured. The thematic information stored in addition to this is gaze-direction. For a detailed description of the .vsp file’s format, see Figure 2.2.
6 CHAPTER 2. BACKGROUND
Figure 2.1: Video footage of campus area
Figure 2.2: The format of the .vsp file.
2.3 Data Visualization
Data visualization is the creation of graphical representations or images of information. It can be viewed as an application of computer graphics, using computer graphics methods to display data. But rather than focusing on purely visual aspects, it aims to aid users in discovering and formulating ideas about datasets through visualization [20].
Consequently it consists of several disciplines such as human-computer inter- action, user perception, statistics and data mining - a combination of computa- tional power and human visual perception[20].
CHAPTER 2. BACKGROUND 7
2.3.1 The Visual Information Seeking Mantra
The Visual Information Seeking mantra - overview first, zoom and filter, then details-on-demand, was first described by Schneiderman in 1996 and is widely cited within information visualization research [3]. It attempts to describe how data is effectively presented to users and often serves as a guideline and as an inspiration for practitioners within the field of data visualization. The mantra consists of six tasks that users of a visualization system should be able to perform:
• Overview first: capture the entire dataset in one view.
• Zoom and filter: remove unnecessary information and reduce the amount of data displayed.
• Details on demand: display additional information if requested, without requiring a change of view.
• Relate: enable the users to observe relationships in the data.
• History: enable users return to a previous state, and compare it to a other states of representation.
• Extract: extract information of interest, so that users do not need to reproduce the same steps of data manipulation to retrieve it again.
In this thesis the mantra has been mainly used as a guideline in the design process, presented in section 3.1.
2.3.2 Techniques to Visualize Trajectory Data
This section will present techniques used to visualize trajectory data, commonly found within geographic information systems.
Static and Animated Maps
Static and animated maps are considered to be the most common methods to visualize movement of discrete objects [2]. These maps often make use of
8 CHAPTER 2. BACKGROUND
linear symbols, such as lines and arrows to represent trajectories. Other visual attributes like color, thickness and transparency can be added to represent events or changes. Static maps are able to capture spatio-temporal information in a specific moment or aggregated over time to display a general picture. However, they are not able to show how data changes over time. An example of such a visualization is shown in Figure 2.4 (a).
Animated maps with an interactive time filter are able to show how relationships in data changes and evolves over time. As in the case of static maps, trajectories are often displayed as linear symbols. The map has been complemented with an interactive slider for time with which the user can select a time interval. With this filter, segments of trajectories present in the chosen time interval are drawn.
In Figure 2.3 and 2.4 (b) examples of such visualizations are demonstrated.
Figure 2.3: Illustration of an animated map. The visualization consists of several maps, each one capturing a moment in time. Source: [12]
CHAPTER 2. BACKGROUND 9
Figure 2.4: (a) A static map over ship routes (b) Animated map with interactive time filter. Source: [2]
2.4 User Evaluation
2.4.1 Usability Evaluation and Common Metrics
ISO 9241-11:2018 states that ”usability focuses on the effectiveness, efficiency and satisfaction of the user’s interaction with the object of interest”, and defines these metrics as the following:
Effectiveness as ”the accuracy and completeness with which users achieve specified goals”, where accuracy is how well the outcome matches the intended outcome, and completeness how well users are able to achieve all intended outcomes.
Efficiency as ”the resources used in relation to the results achieved”, where resources could be time, money, effort or similar.
Satisfaction as ”the extent to which the user’s physical, cognitive and emotional responses that result from use of a system, product or service meet user’s needs and expectations” [1].
10 CHAPTER 2. BACKGROUND
In section 3.2 these metrics are used to test the visualization.
2.4.2 Think Aloud Protocol
A think-aloud prompts the user to speak their thoughts while interacting with a system. These studies typically focus on the interaction between the partici- pants and the system, making them well suited when attempting to identify its strengths and weaknesses [4]. It also offers insight into the cognitive process of the user.
2.4.3 Analytical Tasks for Trajectory Data
When analyzing spatio-temporal data such as trajectories, patterns can emerge in respect to both the past and the present. These patterns consist of changes that occur over time and there are different approaches to categorize them.
Pequet [17], considers spatio-temporal data to consist of three co-dependent parts: space (where), time (when) and objects(what). These parts can be combined into three basic tasks:
• When + where → what: Describe the objects or set of objects that are present at a given location or set of locations at a given time or set of times.
• When + what→ where: Describe the location or set of locations occu- pied by a given object or set of objects at a given time or set of times.
• Where+what→ when: Describe the times or set of times that a given object or set of objects occupied a given location or set of locations.
Adrienko et al [16], extend these questions by adding search levels - elementary and general - when users focus on one or multiple objects. They also argue that cognitive operations of identifying or comparing objects should be considered.
By grouping the parts of when and what + where, the extended task can be presented as four categories:
• Elementary “when” and elementary “what + where”: describe char- acteristics of this object (location) at the given time moment.
CHAPTER 2. BACKGROUND 11
• Elementary “when” and general “what + where”: describe the situa- tion at the given time moment.
• General “when” and elementary “what + where”: describe the dy- namics of characteristics of this object (at this location) over time.
• General “when” and general “what + where”: describe the evolution of the overall situation over time.
How this framework is put to use for the implementation in this thesis is described in further detail in section 3.2.
2.5 Related Work
In this section related works which use relevant visualization techniques are presented. It also features work that have used the framework by Adrienko et al. [16] when designing and validating a product.
2.5.1 Spatio-Temporal Visualizer
In this study by Hugo Sequeira [18] a visualization tool was built using the technique of animated maps. It was developed with web development tools, one of them being D3.js. The objective was to provide users with a tool that enabled them to explore and compare trajectories of objects by observing their movement. The tool utilized a slider to adjust a time interval, making it possible to restrict data with time as a criteria and navigate through it. The dataset used consisted of GPS trajectories, tracking 182 people over a period of 5 years, mainly located in Beijing, China. When designing and evaluating this tool, tasks were constructed using the framework by Adrienko et al [16]. However these tasks were not performed by users in actual usability tests, but rather just to prove that the system could provide answers to questions of this nature.
12 CHAPTER 2. BACKGROUND
Figure 2.5: The spatio-temporal visualizer with time-slider, displaying 24 trajectories over a period of 94 hours. 1
2.5.2 Comparing 2D Static Maps and STC
ST-TrajVis is an application that uses 2D static maps and a space time cube (STC) to visualize trajectory data. The space-time cube is a visualization technique which uses two dimensions to represent space and a third dimension for time. An image of the system with the two techniques side by side can be seen in figure 2.6. A study performed by Gonçalves et al. [13] investigated the usability of the system through a user study with five participants when the system was still in development. The dataset was a subset of the GeoLife project and consisted of trajectories from one user in the time period of one month. Metrics tested were efficiency in terms of time performing tasks and effectiveness in amount of completed tasks and their accuracy. A similar study was made comparing the exact same techniques in [11], but with a larger set of participants. It consisted of 16 people and measured the same metrics.
1Source: https://github.com/hugocore/spatiotemporal-visualizer
CHAPTER 2. BACKGROUND 13
Figure 2.6: The system ST-TrajVis with a 2D map to the left, and a space-time cube to the right. Source: [13]
2.6 Tools
2.6.1 DOM
The Document Object Model is a programming interface supported by most modern browsers. It represents documents, such as HTML or XML docments in an object-oriented manner with a tree structure. A web page is a document and the DOM allows users to manipulate the document. This is can be useful when creating a visualization or a webpage. Often JavaScript is used to add and remove objects from the webpage by manipulating the DOM.[14]
2.6.2 Web-development tools
HTML, CSS and JavaScript are the basic building-blocks of a webpage.
HTML provides the layout of the page, CSS the styling and JavaScript is used to manipulate the DOM.
Bootstrap is a front-end framework that divides the page into a grid-structure, which makes it easy to control the layout of the page [5]. D3.js is a data driven approach to manipulate documents. It helps users to build and manipulate DOM-objects based on the data provided [9].
How these technologies were used in order to create an interactive visualization of the data is described in further detail in section 3.1.
Chapter 3 Method
First, all methods related to building the visualization will be presented - from pre-processing the data to the different parts that the visualization consists of. Secondly, a description of how the user evaluation was conducted will be provided, from constructing user tasks, to how the actual tests were performed.
3.1 The Implementation
The visualization was created using D3.js together with Bootstrap to help with the layout. It consists of three parts - an overview in the form of a heatmap, the animated map as a graph of trajectories, and a slider which provides a tool to filter the data. D3.js was used to read a JSON-file with data and build a path for each trajectory by appending elements to an SVG using D3’s built in functions.
3.1.1 Data Cleaning and Formatting
As illustrated in figure 2.2, the data contained information about gaze direction and lines with control points. This thematic information was irrelevant for the visualization and was therefore removed. A Java-program was built to extract the relevant information and format it as JSON-objects. The format of the resulting JSON objects is as follows:
14
CHAPTER 3. METHOD 15
{ ”person”:0, ”x”:174.0, ”y”:-131.0, ”frame”:0 } These objects where then stored as an array in a JSON-file.
3.1.2 The Visualization
Animated Map
The animated map was the main focus of the visualization, as it provides details about the trajectories. It consists of several paths, each corresponding to a trajectory in the data. Researchers from the Embodied Social Agents Lab (ESAL) at KTH suggested that a head - a symbol like an arrow or circle - should be added to the path. This would indicate the last position of the object and therefore also the direction in which it was moving when animated. To further clarify this, as well as adding an element of time to a static view of the animated map, they also suggested that the opacity of the paths declined the further away from the head the path-segment was. Their motivation behind this design was that one would be able to know if any collisions had occurred and what person visited a certain area first. In figure 3.1 the animated map is illustrated.
Figure 3.1: Left: the animated map displaying 79 filtered trajectories. Right:
the animated map displaying the entire dataset of 415 trajectories.
16 CHAPTER 3. METHOD
Slider
At the top of the page a slider was placed, which allows users to filter the data on frames. The slider was built by using a code example from Mike Bostock found at bl.ocks.org [6], shared under GNU General Public License version 3.
By allowing users to filter on frames they are able explore the data in a small time-interval. When selecting a part of the slider the animated map will update to only show trajectories within the frames that have been selected. This allows users to forward/rewind the time and explore how the trajectories change over time.
Figure 3.2: The slider with a selected interval.
Heatmap
A static heatmap was added in order to provide an overview of the data. This static map can provide an aggregated view of the trajectories over time, illus- trating the density which otherwise has to be computed by the user through observing individual trajectories in the animated map, something which could become a strenuous task with large datasets.
The heatmap consists of 10 000 small rectangles, each corresponding to a specific location in the scene from which the data is taken. Each square has a color corresponding to the amount of people that has been registered at the corresponding location. The heatmap was built by creating a set amount of rectangles and then going through the data, changing the color of rectangles with the same coordinates as the coordinates of the data. The heatmap used in the implementation is provided in figure 3.3.
CHAPTER 3. METHOD 17
Figure 3.3: The heatmap
3.2 Mapping Framework to User Tasks
In order to evaluate the visualization tool, user tasks were created based on the framework in section 2.4.3. An initial literature study produced three instances where this framework was used to construct such tasks. However, only two of these performed actual user tests. Moreover these applications were geographical information systems with datasets containing GPS data [11, 13].
Despite the differences between the implementation in this thesis and previous works, the framework offered a level of abstraction which made it applicable to this project. When constructing the tasks the two levels of elemental and general were taken into consideration - whether users focus on one or multiple objects, as well as the two categories of user objectives identify and compare.
The questions together with their respective category is presented below.
Question 1: Can you identify and give an example where someone is walking at a different speed than someone else?
With this question the user is prompted to compare a difference between two objects - in this case speed. The task belongs to the category ”general when
18 CHAPTER 3. METHOD
and general what+where”.
Question 2: Can you locate any group behaviors in the scene between frame 100 - 1300?
The question asks whether users can identify patterns that occur in crowds - for instance when people form a group and walk together. This task is in the category ”general when and elementary what + where”.
Question 3: Choose a person and track them throughout the scene, can you describe their movement?
This question is also categorized as ”general when and elementary what + where”. However, it is different than the previous question as it tasks users to use the visualization tool by filtering and forwarding/rewinding in time to first identify and then follow a certain individual.
Question 4: Can you tell how many people are present in the scene between the frames 400-900?
This question belongs to the category of ”elementary when and general what+where”
and asks the user to identify the number of trajectories present within a specific interval.
Question 5: Can you tell what the most visited locations are in this scene over the entire time span?
This question is a way to ask users to describe the evolution of the overall situation over time, through identifying hot spots. This is also a task which falls under the category of general when and general what+where.
One category that has not been captured as a question by the work in this thesis is elementary “when” and elementary “what + where”. This is due to the fact that the implementation does not support this kind of interaction yet. It would mean that a specific person could be selected out of the dataset and its characteristics displayed.
CHAPTER 3. METHOD 19
3.3 Pilot
A small pilot study with three participants was conducted, with the primary interest to see how fast users felt familiar with using the time filter and observing animated trajectories. Another objective with the pilot was to test a selection of tasks that had been constructed with the framework suggested by Adrienko et al.
[16] and observe how they translated into action with users. They were asked to think aloud while performing a task and their comments were collected. In particular they were queried on what they thought about following individual trajectories, as well as seeing the overall trend of movement in the scene.
3.4 User Study
After having performed the small scale pilot study, a user test was conducted with more participants. The questions constructed with the framework men- tioned in section 3.2 were used.
The visualization was tested on three different densities scaled linearly from 50 up to 415 trajectories which was the total amount present in the dataset. Thus each test was performed with 50, 230 and 415 trajectories. To rule out any order effects, a total of six participants were recruited and assigned a unique combination of densities to test.
First the participants were informed on the procedure of the test as well as their rights. At the beginning of the experiment, all features of the tool were demonstrated to the participants. A short explanation of the dataset being displayed was given. Each person was given time to try the slider and its effects and ask questions if anything needed to be clarified. Afterwards each participant was asked to perform the five tasks.
Efficiency was not measured, since users were asked to provide their thoughts through the think aloud protocol. This could have affected the completion-time for each task and thus a compromise was made to collect qualitative data.
Chapter 4 Results
4.1 Implementation
In section 3.1.2 a method to create a visualization tool for crowd trajectory data using D3.js is described. The layout of the final product can be seen in 4.1. The separate features of this implementation are shown in figure 4.2 and figure 4.3.
Figure 4.1: The entire visualization tool in one view. Time slider, animated trajectories within selected frame interval and a heatmap.
20
CHAPTER 4. RESULTS 21
Figure 4.2: The animated map and the time slider.
Figure 4.3: The heatmap together with description and legend.
22 CHAPTER 4. RESULTS
4.2 User Test
In the user test participants were asked to perform the five tasks from section 3.2. Effectiveness was measured as completion - whether the task could be completed or not. Satisfaction was measured as ease of use, asking test subjects to rate how difficult it was to perform a task on a scale from 1-5, where 1 was considered easy and 5 difficult.
All users were able to complete all tasks.
A graph of the average difficulty for questions 1-5 for all three densities, as well as well as the average of each individual question can be seen in 4.9. The experienced difficulty for each question is reported as separate graphs below.
Tables of the collected data is also available in Appendix A.
Question 1: Can you identify and give an example where someone is walking at a different speed than someone else?
As seen in figure 4.4, three out of six participants rated the difficulty of this task as a 1 for all densities. There is a notable difference for one participant - rating the difficulty of this particular task at 4 and 4.5, however this user provided no further information to motivate their rating. While performing this task, two participants noted that trajectories with the color of a pale yellow was difficult to distinguish from the rest.
CHAPTER 4. RESULTS 23
Figure 4.4: The experienced difficulty of question 1, for 50, 230 and 415 trajectories for all six participants. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
Question 2: Can you locate any group behaviors in the scene between frame 100 - 1300?
When performing this task all six of the test subjects looked at the trails to see if objects moved in parallel and in a close proximity to one another. All of the test subjects found instances where people were walking in pairs. However, four out six also found a larger group of three people. The experienced difficulty for this task can be seen in figure 4.5.
24 CHAPTER 4. RESULTS
Figure 4.5: The experienced difficulty of question 2, for 50, 230 and 415 trajectories for all six participants. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
Question 3: Choose a person and track them throughout the scene, can you describe their movement?
Users had two different approaches to solve this task. Three users chose a smaller time interval and tracked the movement of the person by moving the interval back and forth, describing the movement step by step. The other three identified where the person entered the scene, selecting this as the beginning of the interval and then continued to expand their selection until noting that an exit had been made. They then had the entire trajectory, with head and tail for the entirety of its presence in the scene to look at and describe. The experienced difficulty for this task can be seen in figure 4.6.
CHAPTER 4. RESULTS 25
Figure 4.6: The experienced difficulty of question 3, for 50, 230 and 415 trajectories for all six participants. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
Question 4: Can you tell how many people are present in the scene between the frames 400-900 ?
The difficulty of performing this task for all three densities can be seen in figure 4.7, where it can be noted that the average experienced difficulty was the same for 50 trajectories as for 230 trajectories. Four out of the six users counted the amount of trajectories instead of reading the number printed at first. However, they found the number eventually while performing the task. Test participant five thought that the number was not printed clearly enough and motivated their rating of difficulty based on this. Two of the users also commented that the slider did not provide enough accuracy and that it was hard to know the exact size of the chosen time interval.
26 CHAPTER 4. RESULTS
Figure 4.7: The experienced difficulty of question 4, for 50, 230 and 415 trajectories for all six participants. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
Question 5: Can you tell what the most visited locations are in this scene over the entire time span?
The experienced difficulty for this task can be seen in figure 4.8. All test subjects used the heatmap to perform this task. Note that the average experienced difficulty for the densities of 230 and 450 trajectories was the same.
One user commented that the heatmap looked too sparse when performing the task at a density of 50 trajectories and therefore rated the difficulty as 2.
Two participants also experimented with the slider to select the entire interval to create an aggregated view of all trajectories - commenting that it was not possible to determine whether or not some trajectories were hidden underneath others offering no aid in the task. However, they also noted that this view provided them with the areas that were not visited.
CHAPTER 4. RESULTS 27
Figure 4.8: The experienced difficulty of question 5, for 50, 230 and 415 trajectories for all six participants. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
The average difficulties for the three different densities can be seen in figure 4.9. On average it appears that question 2 - whether or not any group behaviors could be located - was the most difficult for all three densities. The average difficulty to perform all tasks 1-5 is in the range between 1 and 2, with a small difference between the density of 230 and 415 trajectories.
28 CHAPTER 4. RESULTS
Figure 4.9: The average difficulty for all questions 1-5 tested on 50, 230 and 415 trajectories. Difficulty was rated on a scale from 1 to 5, where 1 was considered easy and 5 difficult.
Chapter 5 Discussion
Techniques from the fields of cartography and geographical information systems were adopted in this project since they offered previous work on visualizations for trajectories and spatio-temporal data. However in this thesis an attempt was made to utilize the techniques to visualize trajectories of crowds rather than that of GPS data from separate individuals. The visualization presented here does not rely on a traditional map, but rather focuses solely on human movement within a limited space - displaying characteristics typical of crowds.
The chosen dataset separates our implementation from previous work, but has a similarity in the fact that it utilizes a time filter and animation - for instance the spatio-temporal visualizer [18], which also was built using D3.js.
The task rated as the most difficult by users was question number two as shown in figure 4.9. This question is asking the user to locate group behaviours in the data - something which occurs in crowds but might not be present in datasets with GPS tracking of separate individuals. It is therefore interesting to investigate further if there are any other features or visualization techniques which might aid in the exploration of group behaviours in trajectory data. This could possibly strengthen the animated map currently used in the implementation.
As for the user tests and creation of tasks, the framework suggested by Adrienko et al. [16] has in this thesis been interpreted as in the previous work of Goncalves et al. [11, 13] such that the levels of elementary and general are directly related to the number of elements involved. Our user test with six participants is comparable in size to [13] where five participants tested a system still in development. We think that despite there being a small number of participants
29
30 CHAPTER 5. DISCUSSION
used in this work - it was able to yield results that illustrated interesting problems and possible improvements for the visualization. However at a later stage of development we believe it would be desirable to increase the amount of test subjects to gather more quantitative data to analyze. It is also important to mention that although each participant got time to familiarize themselves with the tool it is still not possible to completely rule out learning effects, as users could have felt more comfortable with the tool the longer they interacted with it.
For the effectiveness measure, all users were able to complete all tasks. This is a positive result, but for another iteration of user tests it would be of interest to go even further with this metric and test a level of accuracy. This could be done by measuring the accuracy of the answers provided by users for each task. As for the framework, we have been able to show that it is applicable to a visualization utilizing a dataset with crowds, despite its differences from datasets in previous works [18, 11, 13].
Since the average experienced difficulty for all five questions is in a range between 1 and 2, as seen in 4.9, for the maximum amount of trajectories available in this dataset - there could still be room to test how the system scales with an even larger set of data.
As previously mentioned problems and suggested improvements were brought to attention by the user test. One of these problems was that one of the colors of the trajectories - a pale yellow, was difficult to distinguish from the rest. This can easily be solved by using a darker color. It was also mentioned that the slider for time could be improved by complementing it with an input field for the end points of an interval. By adding this feature one can keep track of the interval’s size as well as the start and end of the interval with greater accuracy, compared to the drag and click option of the slider. We hypothesize that the issue with users not noticing that the number of trajectories was printed, could be due to the fact that the text size was small and placed beneath a static header.
Another issue to address is the fact that trajectory data could be sparse in the amount of captured points of each trajectory. Since the heatmap is built on these captured points, this could result in sparse areas in the heatmap despite it being visited by pedestrians. We therefore advice that the data is interpolated in order to avoid any potential problem.
CHAPTER 5. DISCUSSION 31
5.1 Future Work
Considering that the implementation does not support tasks of the category elementary ”when” and elementary ”what+where” this could be a natural next step in the development of the visualization. The feedback provided by the users in the think aloud also offers insight into things which should be further improved with the implementation, as mentioned in the previous section.
As for the user tests, the amount of participants could be made larger at a later stage of development. It would also be interesting to introduce an aspect of accuracy to the effectiveness metric, as well as measuring efficiency.
Chapter 6 Conclusion
We have created a visualization consisting of an animated map, a heatmap and a slider for filtering the data. The evaluation shows that this approach is applicable on small datasets of crowd trajectories - but it also implies that it could be possible to test even larger datasets than what has been done within the scope of this thesis.
We can conclude that our visualization is a tool that can aid a user in analyzing and finding patterns in pedestrian trajectory data.
We have also been able contribute to the collection of works [18, 11, 13] that have implemented the framework suggested by Adrienko et al. [16] to evaluate visualizations of human trajectory data. Since the work in this thesis concerns crowds, this could provide a reference for future work attempting to evaluate a system with data of this nature.
32
Bibliography
[1] ISO 924111:2018(E). Ergonomics of human-system interaction —Part 11: Usability: Definitions and concepts. Standard. Geneva, CH: Interna-
tional Organization for Standardization, Mar. 2018.
[2] Natalia Andrienko and Gennady Andrienko. “Visual analytics of move- ment: An overview of methods, tools and procedures”. In: Information Vi- sualization 12.1 (2013), pp. 3–24. doi: 10.1177/1473871612457601.
url: https://doi.org/10.1177/1473871612457601.
[3] P. Cairns B. Craft. “Beyond guidelines: what can we learn from the visual information seeking mantra?” In: Information Visualisation, 2005.
Proceedings. Ninth International Conference on (2005).
[4] Ann Blandford, Dominic Furniss, and Stephann Makri. “Qualitative HCI Research: Going Behind the Scenes”. In: Synthesis Lectures on Human-Centered Informatics 9.1 (2016), pp. 1–115. doi: 10.2200/
S00706ED1V01Y201602HCI034. eprint: https://doi.org/
10 . 2200 / S00706ED1V01Y201602HCI034. url: https : / / doi.org/10.2200/S00706ED1V01Y201602HCI034.
[5] Bootstrap. https://getbootstrap.com/. Accessed: 2018-05- 12.
[6] Brush snapping. https://bl.ocks.org/mbostock/6232537.
Accessed: 2018-05-13.
[7] Rose Challenger, Chris W. Clegg, and Mark A. Robinson. “Emergency Planning College Understanding Crowd Behaviours: Guidance and Lessons Identified”. In: (June 2009).
[8] C. Chen. “Top 10 unsolved information visualization problems”. In:
IEEE Computer Graphics and Applications 25.4 (July 2005), pp. 12–16.
issn: 0272-1716. doi: 10.1109/MCG.2005.91.
33
34 BIBLIOGRAPHY
[9] D3.js. https://d3js.org/. Accessed: 2018-02-16.
[10] Niklas Elmqvist and Ji Soo Yi. “Patterns for visualization evaluation”. In:
Information Visualization 14.3 (2015), pp. 250–269. doi: 10.1177/
1473871613513228.
[11] Tiago Gonçalves, Ana Paula Afonso, and Bruno Martins. “Cartographic visualization of human trajectory data: overview and analysis”. In: Jour- nal of Location Based Services 9.2 (2015), pp. 138–166. doi: 10.1080/
17489725.2015.1074736.
[12] Tiago Gonçalves, Ana Paula Afonso, and Bruno Martins. “Visualization techniques of trajectory data: Challenges and limitations”. In: CEUR Workshop Proceedings 1136 (2013).
[13] Tiago Gonçalves et al. “ST-TrajVis: Interacting with Trajectory Data”. In:
Proceedings of the 27th International BCS Human Computer Interaction Conference (HCI 2013) (2013).
[14] Introduction to the DOM. https://developer.mozilla.org/
en-US/docs/Web/API/Document_Object_Model/Introduction.
Accessed: 2018-05-14.
[15] Krumm J. Lee WC. “Trajectory Preprocessing”. In: Computing with Spatial Trajectories. Springer, 2011. Chap. 1, pp. 3–33.
[16] P. Gatalsky N. Adrienko G. Adrienko. “Exploratory spatio-temporal visualization: an analytical review”. In: Journal of Visual Languages &
Computing 14 (2003), pp. 503–541.
[17] Donna J. Peuquet. “It’s about Time: A Conceptual Framework for the Representation of Temporal Dynamics in Geographic Information Sys- tems”. In: Annals of the Association of American Geographers 84.3 (1994), pp. 441–461. issn: 00045608, 14678306. url: http://www.
jstor.org/stable/2563777.
[18] Hugo Sequeira. “Spatio-Temporal Visualizer: Online tool to visualize spatiotemporal data with an interactive time-window slider, using D3.js, Leaflet.js and Crossfilter”. In: (July 2014).
[19] Anna-Katharina Lautenschütz Somayeh Dodge Robert Weibel. “Towards a taxonomy of movement patterns”. In: Information Visualization 7.3-4 (2008), pp. 240–252. doi: https://doi-org.focus.lib.kth.
se/10.1057/PALGRAVE.IVS.9500182.
BIBLIOGRAPHY 35
[20] Chad A. Steed. “Interactive Data Visualization”. In: Data Analytics for Intelligent Transportation Systems. Ed. by Jan Fagerberg, David C. Mowery, and Richard R. Nelson. The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom: Elsevier, 2017. Chap. 7, pp. 165–190.
[21] M. Tory T. Moller. “Human factors in visualization research”. In: IEEE Transactions on Visualization and Computer Graphics 10 (2004), pp. 72–
84. doi: 10.1109/TVCG.2004.1260759.
[22] UCY Computer Graphics Lab. https://graphics.cs.ucy.ac.
cy/research/downloads/crowd- data. Accessed: 2018-02- 16.
Appendix A
Results in Tables
Density: 50, avg. difficulty: 1.37
Participant Question 1 Question 2 Question 3 Question 4 Question 5
1 1 1 1 1 1
2 1 1 1 1 1
3 1 1 1 1 1
4 4 2 1 1 1
5 1 2 2 3 1
6 1 3 1 1 2
Density: 230, avg. difficulty: 1.52
Participant Question 1 Question 2 Question 3 Question 4 Question 5
1 1 1 2 1 1
2 1 1 1 1 1
3 2.5 3 2 1 1
4 4 2 1 1 1
5 1 3 2 3 1
6 1 2 1 1 1
36
APPENDIX A. RESULTS IN TABLES 37
Density: 415, avg. difficulty: 1.58
Participant Question 1 Question 2 Question 3 Question 4 Question 5
1 1 1 2 1 1
2 1 1 2 2 1
3 1 4 2 1 1
4 4.5 2 1 1 1
5 1 2 1 4 1
6 1 2 2 1 1
www.kth.se
