Driver Attention Allocation for Information Acquisition

行政院國家科學委員會

獎勵人文與社會科學領域博士候選人撰寫博士論文

成果報告

Driver Attention Allocation for Information

Acquisition

核 定 編 號 ： NSC 100-2420-H-009-003-DR 獎 勵 期 間 ： 100 年 08 月 01 日至 101 年 07 月 31 日 執 行 單 位 ： 國立交通大學交通運輸研究所 指 導 教 授 ： 汪進財 博 士 生 ： 黃士軒 公 開 資 訊 ： 本計畫可公開查詢

中 華 民 國 102 年 10 月 03 日

國 立 交 通 大 學

交通運輸研究所

博士論文

No.075

駕駛注意力分配模式

A Novel Approach

for Modeling Driver Attention Allocation

指導教授：汪進財

研

究 生：黃士軒

駕駛注意力分配模式

A

N

OVEL

A

PPROACH FOR

M

ODELING

D

RIVER

A

TTENTION

A

LLOCATION

研 究 生：黃士軒 Student：Shih-Hsuan Huang

指導教授：汪進財 博士 Advisor：Dr. Jinn-Tsai Wong

國 立 交 通 大 學

交通運輸研究所

博 士 論 文

A Dissertation

Submitted to Institute of Traffic and Transportation

College of Management

National Chiao Tung University

in partial Fulfillment of the Requirements

for the Degree of Doctor of Philosophy

in

Management

July 2013

Taipei, Taiwan, Republic of China

駕駛注意力分配模式

學生：黃士軒

指導教授：汪進財教授

國立交通大學 交通運輸研究所

摘要

注意力分配為行車安全的重要關鍵，駕駛人必須將其有限資源妥適分配至車輛前方、車內與 車外等區域，以維持適當之情境察覺，並得與前車保持安全距離。然而過去對於注意力分配之研 究多侷限於對各焦點的總體分析，無法探究其個體行為特性，此外，對於視線移轉過程之呈現方 式往往過於著重於「前方」焦點，致使研究成果當中的多數路徑皆為移往或來自前方，此一現象 導致研究無法完整呈現駕駛人視線移轉的完整過程，因此，如何呈現注意力分配以及分析其特性 為事故分析與防範的重要基礎，唯有提出適切的量化方法，才能正確呈現駕駛人移轉注意力之過 程。 以注意力分配為題，本研究欲回答下列問題：1) 駕駛人注意力分配應如何呈現？ 2) 駕駛人 是否會採取特定的注意力分配型態？ 3) 若有，有哪些型態？ 4) 有哪些變數會影響駕駛人注意 力分配？為回答上述問題，本研究首先提出「注意力分配循環」之概念，以前方焦點視為一基準 點，將視線移開前方至他處最後再回到前方的循環過程視為注意力分配之基本元件，呈現駕駛視 線分配的完整過程。本研究於分析階段採用美國 100-car 自然駕駛資料庫當中的事件資料庫進行分 析，透過序列關聯法則，找出駕駛人將視線在前方與非前方之間移轉的路徑，並透過羅吉特模式 之應用，計算駕駛人在不同狀況下，選擇不同類型視線移轉與選擇各焦點的機率分配。 研究結果發現，研究發現超過 90%以上的注意力分配循環僅包含一個非前方焦點，亦即當駕 駛人將視線移開時，多數僅會注視一個非前方焦點 (以車內分心、左後視鏡與車內後視鏡為主)， 以避免因移開視線時間過長而無法觀察前方路況之狀況發生；本研究亦發現駕駛人會將視線在前 方與同一焦點間來回移轉，此特性尤以車內後視鏡與車內分心最為明顯，顯示駕駛人於該二焦點 收集資訊時，會不斷將視線移回前方，以確保將視線移開前方的過程中仍可維持對前方的情境察 覺能力。此外，駕駛人會避免將視線直接自一非前方焦點轉移至另一非前方焦點，而是先將視線 移回前方再移往下一焦點，以確保車前安全。駕駛人選擇焦點時，傾向將視線分配至較近、對安 全影響較大、較明亮且資訊出現頻率較高的焦點上。 最後，本研究將注意力分配循環之概念應用於安全評估，並以駕駛反應時間為基礎，設定駕 駛人得以將視線移開前方的最長時間。研究發現，當駕駛人連續注視的焦點數越多時，其無法觀 察前方路況的總時間越長，因此，對前方刺激的餘裕反應時間越短；其中，分心、駕駛操作意向 等因素皆會影響其實際的反應時間長度。若以注意力分配角度出發檢視，目前現行之 2.5 秒反應 時間設計標準已無法滿足現況，若以 90 百分位為基準，道路設計應將反應時間設定為 3 秒。 囿於資料限制，本研究所引用之資料雖無法代表駕駛人的典型注意力分配型態，然而所提模 式與其結果仍可提供後續分析探討之參考，並可作為事故防範與安全分析之用，此一領域仍待後 續進一步探討。 關鍵字：注意力分配、循環、視線移轉、分心、自然駕駛

A

N

OVEL

A

PPROACH FOR

Modeling

D

RIVER

A

TTENTION

A

LLOCATION

Student：Shih-Hsuan Huang Advisors：Dr. Jinn-Tsai Wong

Institute of Traffic and Transportation

National Chiao Tung University

A

BSTRACT

Attention allocation is the key of driving safety, which relies on the adequate distribution of the driver’s attention to the forward area and to other non-forward focal points. However, most representation of attention allocation are the aggregated result of vision transition. It is not able to observe drivers’ microscopic behavior against dynamic changing environment. Moreover, thus far, current methods seem to be over-emphasized on the dominant forward area, causing the observed paths were mostly ones shifting from or heading to the frontal side. The whole process of transiting vision among focal points cannot be observed. Consequently, a mechanism for attention allocation is a critical issue in crash prevention.

There are four questions that this study aims to answer: 1) How is driver attention allocation represented? 2) Do patterns of driver attention allocation exist? 3) If yes, what are these patterns? and 4) What are the factors affecting driver attention allocation? To answer these questions, this study proposes the concept of renewal cycle, which is the entire process of drivers glancing at forward side, transiting vision away, and finally transiting vision back to the front. Using the renewal cycle as the basic component of attention allocation analysis, this study is able to represent drivers’ vision transition in a more realistic way. In the section of empirical analysis, this study adopted the event database of 100-car naturalistic driving studies. Sequential rule mining and multinomial logit model were used for generating the patterns and probability of drivers transiting vision among focal points.

This study found that over 90% of drivers’ attention allocations were 2-glance renewal cycles, suggesting that drivers usually glance only one off-road focal point, among which the in-vehicle distraction, left mirror and rearview mirror are the three most frequent appeared ones. Among these 2-glance renewal cycles, some were found repeatedly appeared several times, particularly the ones related to in-vehicle distraction and rearview mirror. It suggests a compensation of lost awareness against leading area by separating their long glance off-road into several shorter ones. In addition, drivers prefer not to transit vision from one non-forward focal point directly to another. Instead, they glance at forward side between two non-forward glance for checking the timely status ahead. As for the choices of focal points, four constructs of attributes (Salience, effort, expectancy and value) in SEEV model were included in this model. The result shows that drivers would allocate more attention to the focal point with higher information expectancy and value. On the other hand, less salient and higher effort would inhibit the vision transition.

Finally, this study adopted the Perception Reaction Time (PRT) as the reference for setting the maximum time for drivers to transit vision away from the frontal side. It clearly indicated that drivers glancing consecutively at more non-forward focal points in a sequence were more likely to have insufficient time for responding to harmful changes in front of them. In addition to distractions,

maneuver intentions, number of glances in a renewal cycle, and their interactions all significantly affected drivers’ attention allocation. As for the current 2.5-s PRT rule, it may not be robust enough to satisfy every situation. Based on the results derived from the 100-car event database, a 3.0-s PRT may be better for designing safer roads.

Although the sample drivers adopted in this study were not representative, the preliminary research results were promising and fruitful for potential applications, particularly educating novice drivers. These findings might have striking implications for accident prevention. This area of study deserves further attention.

謝誌

畢業的氛圍總是會讓人回想起過去，八年前推甄沒能錄取成大，當時專題指 導老師戴佐敏博士說：「你去台北找汪老師吧！」，就這樣一句話，我選擇了交大 交研所，研究主題從天上飛變成地上爬，碩士班畢業後不小心念了博士班，這一 待又是六個年頭。彷彿就像起士理論一樣，巧合環環相扣，成就了當初從來不曾 預期的現在。這一路上峰迴路轉，碰到很多貴人與機緣，所以值得很多很多感謝。 陪公子念書，一念就是八年，感謝汪進財教授，學生不才笨鳥慢飛，跌跌撞 撞一段時間，也真的讓您失望了好些日子，感謝您始終耐著性子，一再給學生機 會成長、學習，您的風範，讓我即使被罵了，心頭還是暖暖的。在口試階段，承 蒙周榮昌教授、陳莞蕙教授、曾平毅教授與鍾易詩教授的細心斧正與寶貴意見， 使論文更臻完備。感謝黃承傳教授、黃台生教授、藍武王教授、馮正民教授、陳 穆臻教授與邱裕鈞教授在課堂與論文研討的指導，以及洪姐、柳姐與何姐在行政 事務上的協助，感謝您們為學生創造了如家一般的求學環境。感謝吳昆峰博士在 資料收集的協助，讓我柳暗花明又一村。感謝成大交管系戴佐敏教授，引領我進 入學術殿堂，即使學生離開成大也還是持續關心照顧。感謝運輸學刊工作團隊， 與各位合作這幾年，學生獲益良多。感謝北科大互動所李來春教授，雖然只與老 師相處一個學期，平心而論，那是我博士生涯一個重要的轉捩點。 從研究室菜鳥到老鳥，我何其有幸能與諸位博士班成員一起共事，感謝文健、 昱凱、孟佑、世昌、彥蘅、志誠、日新、沛儒、昭弘、永祥、群明、文斌、立欽、 昭榮、素如、季倫、啟椿等諸位學長姐的照顧，感謝同屆同學熙仁、彥斐、嘉陽、 聖章、姿慧，這是條長遠的路，希望大家都能走到終點，另外還要感謝裕文、季 森、志偉、郁珍、俊宇、奚若、佳欣、鄒亞等學弟妹，祝福各位投稿順利早早畢 業。感謝舒馨口琴 34 屆，一晃眼我們認識 15 年了；感謝 FCS 94，希望吃喝玩 樂團能夠歷久彌新；感謝宇函與阿凱，謝謝你們這些年的支持，有你們真好。 最後的感謝要留給我最愛的家人，感謝一直以來支持我的爸媽，謝謝您們為 我們兄弟倆的付出，辛苦了！感謝自小最疼我的奶奶，這本論文要獻給您，與在 天上保佑我們的爺爺。感謝婷雅，陪伴我度過這段蟄伏，希望未來，我們能繼續 牽著手，一步一步往前走。 感謝陳之藩博士，您那句「要感謝的人太多了，所以就謝天吧！」實在是太 好用了！感謝過去、現在、未來的每一位你們，感謝老天讓我們相遇，引領我走 向每個緣分，好的壞的，人生就是這樣才完美。

黃士軒

筆於交研所最後一夜 2013.7.31

T

ABLE OF

C

ONTENTS

摘要... I

ABSTRACT ... III

誌謝...V

TABLE OF CONTENTS ... VII

LIST OF FIGURES ... IX

LIST OF TABLES ... XI

CHAPTER 1 INTRODUCTION ... 1

1.1 Background and Motivation ... 1

1.2 Research Problems ... 4

1.3 Research Objectives ... 4

1.4 Research Scope ... 5

1.5 Research Flow Chart ... 6

CHAPTER 2 ESSENCE OF DRIVING SAFETY ANALYSES ... 9

2.1 Crash Pattern Analysis ... 11

2.2 Driver Behavior Analysis ... 12

2.3 Attention Allocation ... 14

2.3.1 Definition ... 14

2.3.2 Conceptual framework ... 16

2.3.3 Representations ... 18

2.3.4 Contributing factors ... 19

2.4 Role of Intelligent Transpiration System (ITS) in Attention Allocation .. 22

2.5 Summary ... 26

CHAPTER 3 MICROSCOPIC MODEL... 29

3.1 Model Concept ... 29

3.1.1 Vehicle drivers’ domain ... 29

3.1.2 Methodology: Multinomial logt model (MNL) ... 35

3.1.3 Model specification ... 37

3.2 Numerical Data ... 38

3.3 Model Estimation ... 41

3.4 Discussion ... 43

CHAPTER 4 RENEWAL CYCLE FOR ATTENTION ALLOCATION ANALYSES ... 45

4.1 Definition of Renewal Cycle ... 45

4.2 100-car Naturalistic Driving Data ... 46

4.3 Attention Allocation Analysis from a Renewal Cycle Approach ... 48

4.3.1 Research framework ... 48

4.3.3 Distribution of renewal cycle under varying conditions ... 50

4.3.4 Repeated renewal cycle ... 53

4.4 Pattern analysis ... 55

4.4.1 Methodology: Sequential association rule mining ... 55

4.4.2 Pattern generation ... 57

4.5 Modeling Attention Allocation ... 59

4.5.1 Model specification ... 59

4.5.2 Model estimation ... 63

4.5.3 Characteristics of vision transition... 68

4.5.4 Off-road glances with multiple focal points ... 69

CHAPTER 5 ROAD SAFETY FROM AN ATTENTION ALLOCATION PERSPECTIVE ... 71

5.1 Perception Reaction Time (PRT) ... 71

5.2 Analysis Procedure ... 72

5.3 Duration Analysis... 73

5.4 Desired PRT from Attention Allocation Perspective ... 80

CHAPTER 6 CONCLUSION AND RECOMMENDATION ... 83

6.1 Conclusion ... 83

6.2 Policy Implications ... 88

6.3 Recommendations ... 90

REFERENCES ... 93

L

IST OF

F

IGURES

Figure 1-1 Research flow chart ... 7

Figure 2-1 Comprehensive framework of crash analysis ... 10

Figure 2-2 Model of divided attention ... 15

Figure 2-3 Process of driving attention allocation ... 16

Figure 2-4 The four components of the SEEV model ... 20

Figure 3-1 Concept of vehicle driver’s domain ... 30

Figure 3-2 Important factors of perception domain ... 31

Figure 3-3 Important factors of critical domain ... 32

Figure 3-4 Important factors of reaction domain ... 33

Figure 3-5 Data generation procedure ... 39

Figure 3-6 Hypothetical scan paths... 40

Figure 3-7 Statistics of a hypothesis data set of drivers’ glances... 41

Figure 4-1 Research framework for attention-allocation analysis ... 48

Figure 4-2 Distribution of time between non-forward glances in repeated renewal cycles ... 55

Figure 4-3 Major types of vision transitions ... 60

Figure 4-4 Framework of vision transition model ... 60

Figure 4-5 Probability of directly transiting vision to other non-forward focal point . 69 Figure 5-1 Cumulated probability of glance duration... 73

Figure 5-2 Cumulated probability of glance duration at non-forward focal points under different maneuver ... 76

Figure 5-3 Cumulated probability of glance duration at non-forward focal points with and without distractions ... 77

Figure 5-4 Cumulated probability of glance durations at non-forward focal points under different levels of traffic density ... 79

L

IST OF

T

ABLES

Table 3-1 Parameters for data generation ... 40

Table 3-2 Estimation results of attention allocation model ... 42

Table 3-3 Focal point transition matrix for different levels of glance duration (%) .... 43

Table 4-1 Attributes of 100-car event database ... 47

Table 4-2 Number of glances in renewal cycles ... 49

Table 4-3 Distribution of renewal cycles by attributes ... 51

Table 4-4 Glance duration for non-forward focal points ... 53

Table 4-5 Definition of sequence, element and event ... 56

Table 4-6 Attention allocation patterns of various maneuver intentions ... 57

Table 4-7 Specification of multinomial logit model ... 61

Table 4-8 Estimated logit models for the path from the left side ... 64

Table 4-9 Estimated logit models for the path from the right side ... 65

Table 4-10 Estimated logit models for the path from rearview mirror ... 66

Table 4-11 Estimated logit model for path from in-vehicle distraction ... 67

Table 5-1 ANOVA of glance duration under different maneuver intentions. ... 74

Table 5-2 ANOVA of glance duration under different distraction conditions ... 76

Table 5-3 ANOVA of glance duration under different Levels of Service (LOS) ... 78

Table 5-4 ANOVA of glance duration at different times of day ... 79

Table 5-5 ANOVA of glance duration under different weather conditions ... 80

CHAPTER 1

I

NTRODUCTION

1.1 Background and Motivation

Crash predictability has long been a controversial issue. Bortkiewicz, usually considered the pioneer of modern crash research, stated that crash occurrences are random and thus inexplicable in his 1898 study (Elvik 2006). However, the development of modern analysis techniques has inspired various attempts to explore the causality of accidents. In recent days, it is suggested that scenario of crash-proneness do exist (Visser et al. 2007). Exploring the causes of motor vehicle crashes has become a pressing issue. Finding the causality of crashes is thus possibly one of the most effective ways to improve road safety and to prevent crashes from happening.

To enhance understanding of crashes, researchers have worked on mining aggregated crash data to extract crash patterns. Numerous contributing factors have been found critical to roadway safety. For example, rear-end accidents increased with the number of signal phases and width of traffic island (Chin and Quddus 2003). Demographic characteristics such as age and gender also have been extensively studied (Clarke et al. 1998, Clarke et al. 1999, Chang and Yeh 2007). Despite the significant effect of single factors, recent research has claimed that crashes should be analyzed from a chain perspective (Elvik 2003, Wong and Chung 2007b, Verschuur and Hurts 2008, Wong and Chung 2008a, Wong and Chung 2008b). In addition to the scenario of crash occurrence, some remote factors of crash occurrence must be considered. For example, personality traits can be treated as prior-to-driving factors that affect risky driving behavior (Wong et al. 2010b, Wong et al. 2010c).

Exploring accident chains provides valuable clues that indicate accident-prone scenarios in which drivers usually have a higher risk of being involved in a dangerous situation. However, a crash-proneness driver driving in a crash-proneness scenario does not necessarily lead to the occurrence of crashes. Such accident-prone scenarios explain mostly the conditions in which drivers face higher risks of being involved in a crash, and possibly the mechanism through which such crashes occur. For example, Wong and Chung (2007b) found that young and inexperienced student drivers had an increased likelihood of being involved in off-road accidents on roads with speed limits between 51 and 79 kph under normal road conditions. The reality is that for each accident under certain conditions, there are numerous young and inexperienced student drivers who drive under identical conditions without experiencing accidents. In fact, the majority of crashes are considered preventable, provided that the

surrounding area is properly observed by the driver and adequate maneuvers are successfully executed (Wong et al. 2010b).

It is clear that there is a missing link between crash proneness scenarios and the crash occurrences. Knowing the causality of crashes behind accident chains is the most crucial element in crash analysis and prevention. In fact, different drivers react differently in identical situations. While most drivers can still drive safely in a high accident risk scenario, but some fail to maintain safety, resulting in dangerous situations. Extraction of such crash pattern and possible crash-proneness driving population can only reveal partial nature of crash occurrence. The question thus remains: How do different reactions to identical conditions result in various outcomes. Answers to the question rely on the understanding of drivers – the decision-maker of a running vehicle.

Research conducted in various countries has suggested that the human factor is the most important contributor to crash occurrence. Among those human factors, misallocating attention is one of the most critical cause of crashes or near-crash circumstances (Brown et al. 2000, McKnight and McKnight 2003, Underwood et al. 2003a, Underwood et al. 2003b, Chen et al. 2005, Dahlen et al. 2005, Underwood 2007, Di Stasi et al. 2009, Olson et al. 2009, Chan et al. 2010). In Taiwan, drivers failing to note roadway conditions accounted for 17% of the fatal crashes in 2011 (MOTC 2012). Presumably, a failure to allocate attention appropriately can be seen as the missing link between crash-prone scenario and crash occurrence within an accident chain. Problems of dividing limited attention resource would cause longer reaction time and higher crash possibility (Cheng et al. 2011). Thus, understanding the patterns of attention allocation is crucial to analyzing the relationship between crashes and ways to maintain situational awareness through visual transition.

Safe driving requires drivers to pay continued attention to various areas and to constantly update awareness of the driving environment. Information perception, which is the first stage of Ensley’s situational awareness, is the key step of comprehending, anticipating, and reacting against tasks or events (Endsley 1995). Acquisition of incomplete or useless information will lead to insufficient comprehension of the current driving environment, misjudgment, rush reaction, and possibly to a crash. To drive safely, drivers must pay attention to multiple sources of information to make informed driving decisions. However, one’s mental resources are limited (Kahnemen 1973). Each driver has a central processor that determines the policy of attention allocation, which divides their mental resources within the limits of their mental capacity. Problems of divided attention may degrade one’s ability to detect potential threats while driving (de Waard et al. 2008, de Waard et al. 2009,

Marmeleira et al. 2009). Complex driving tasks with more information that drivers must attend to would cause drivers making more errors (Elvik 2006).

Distraction is one of major causes of attention misallocation. Shifting attention away from driving to undertake secondary tasks, such as answering cell phones, may increase the time required to perceive and react to external stimuli, and, thus, increase the risk of crashes (Neyens and Boyle 2007a, 2008). Providing drivers with information via in-vehicle information systems, such as GPS, is intended to help drivers more effectively plan the allocation of mental resources and prevent dangers from uncertainty. However, improper use of such devices can yield a negative effect and cause drivers to miss critical information (Liang et al. 2007, Wong and Chung 2007b, Vashitz et al. 2008). Horrey and Wickens (2007), a driving simulation experiment, stated that long glances over 1.6 seconds inside vehicles accounted for 86 percent of crashes. Klauer et al. (2006) also stated that shifting vision away from the forward area longer than 2 s increases the crash/near-crash risk by at least twofold.

It is obvious that a malfunction of attention allocation is the critical link that connects crash-prone scenario with crash occurrence. Misallocating attention may result in one’s awareness being distracted by useless information; thereby missing important information. In just a fraction of a second, one’s visual inattention can lead to unsuccessful information perception. Maneuvering without sufficient information of road conditions could generate unsafe situations easily and increases the likelihood of driver error. To explore the causality of crashes and to prevent them from happening, a functional mechanism for attention allocation is a vital issue that should be tackled. Knowledge of the patterns in which drivers allocate attention among multiple focal points provides insight into the information-seeking behavior of drivers and its relationship to safety.

Unlike those measurable attributes (such as roadway, environment or maneuver conditions) used in crash causation analyses, exploring attention allocation mechanism may face difficulties of observing a driver’s inherent behavior. Fortunately, the recent technique improvement enables the large scale data collection, including eye movement, bio-medical signal and associated maneuvering behavior. For example, the naturalistic driving studies were widely conducted for recording drivers’ every motion of attention allocation and maneuvering. Such a method provides ample opportunities for researchers to further explore drivers’ characteristics from mental and cognitive perspectives. Grabbing those chances would help explore the accident chain in deeper depth and bridge the missing link between crash occurrence and crash-proneness scenarios.

1.2 Research Problems

Demonstrating a driver’s behavior of attention allocation is a challenging issue in various aspects. Mental model is a complex system which contains numberless rules for driver to allocate attention, perceive information, and take actions against dynamic driving tasks. A sophisticated model of attention allocation must be able to reflect the distinct pattern that drivers shift attention between potential sources of driving information.

The first and the most fundamental problem that this study must solve is the representation of attention allocation. Driver attention is not a manifest variable that can be measured directly. Thus, developing an appropriate representation of attention allocation is challenging. An adequate attention allocation representation should enable representing the continuous process of drivers transiting attention from one area to another, and allow researchers to examine the characteristics of different focal points in naturalistic driving tasks. Following the development of an attention allocation representation, the core process of attention allocation is the allocation mechanism, which determines one’s decisions in selecting a specific target for observation. One question must be asked: do driver have an explicit pattern to allocate attention? If yes, what are these patterns? Finally, focal points do not attract drivers’ attention randomly. Some cues from environmental conditions, traffic flow and roadway devices may direct drivers attention in distinct ways. Finding the factors and examining the way they affect attention allocation is a serious issue for identifying the potential risk-proneness sites.

All in all, this research is trying to explore attention allocation by examining the following problems.

(1) How is driver attention allocation represented?

(2) Do patterns of driver attention allocation exist?

(3) If yes, what are these patterns?

(4) What are the factors affecting driver attention allocation?

1.3 Research Objectives

The objectives of this research are two-fold.

In this study, the representation aimed to quantify the unobservable attention for analyzing its characteristics, and to analyze its relation to driving behavior. Different representation method may reveal varying aspect of attention and play essential role in interpreting situational awareness strategies. This study aims to explore the paths of drivers transiting vision from one focal point to another. Based on the path approach that has been utilized, this study proposes a new representation of Renewal Cycle to reflect drivers’ naturalistic driving behavior.

(2) Identify the patterns of attention allocation that drivers commonly held under varying conditions:

The primary goal of this study is identify whether there is a pattern that drivers usually hold to transit vision. If the pattern do exist, this study should be able to explore and represent the drivers’ central mechanism of governing their attention allocation. To reach the goal, the study reviewed previous research for identifying the factors contributing to attention demand of a focal point. Then, based on the contributing factors and the representation proposed, the process of attention allocation was analyzed for deriving its characteristics and the scan path of vision transition while driving. Moreover, it has been stated that the driving safety relies on observing every individual motion that drivers make against driving tasks (Laureshyn et al. 2010). Thus, a microscopic model of attention allocation was estimated for deriving the probability of choosing specific focal point. In this model, the choices of different type of vision transition and the path of transiting vision among focal points were analyzed and presented.

(3) Incorporate the contributing factors that may affect the attention demand of a focal point and vary the drivers’ vision transition process:

Driving in a dynamically changing environment. Numerous factors would vary drivers’ attention allocation in different ways. One of the objective of this study is to select the contributing factors based on literature review and to include the factors into models for evaluating their effect on attention allocation.

1.4 Research Scope

(1) Only visual attention was included.

information using different senses, such as sense of sight, hearing or touch. Seeing that the visual stimuli accounted for the majority part of driving information (Ho 2008, Shinar 2008), this study considered only the visual attention and treated the visual glance to focal point as attending to gather information.

(2) “Looked but failed to see” was not included.

Consciousness and attention are two similar but distinct concepts. Sometimes, drivers may allocate their attention and direct vision to a selected target. Yet, in the level of consciousness, attributes of the targeted object are neither identified nor perceived. However, due to limitation of 100-car dataset, the phenomenon of “Looked-but-Failed-to-See” was not discussed. This study did not differentiate if drivers consciously perceive the information they intended to gather.

1.5 Research Flow Chart

Aiming to answer the problem and reach the goal of this research, this research was organized as Figure 1-1. Noting that attention allocation is critical in perceiving information and making decisions, clarifying the connection between attention allocation and accident chain can help explore the essence of crashes. In Chapter 2, the literatures regarding the crash analysis and attention allocation were reviewed. Particularly, the factors affecting attention demand were discussed. On the basis of these works, the framework of a microscopic driver attention allocation model was proposed. Prior to the validation process, a numerical study is performed to identify the feasibility and appropriateness of proposed model. Advantages and limitations of this model were discussed. Then, in Chapter 4, the concept of Renewal Cycle was proposed. Using the 100-car naturalistic driving data, the attention allocation process was analyzed and modeled. Safety performance was evaluated based on this concept. In Chapter 5, this study evaluated the safety performance from the renewal cycle perspective. Finally, the model applications in driving safety and the conclusions were made in Chapter 6.

Essence of Roadway Safety Analysis

Driving Behavior Information System for

Safety

Attention Allocation Driving Environment

Concept of Microscopic Attention Allocation Model

Numerical Dataset

Concept of Renewal Cycle

Characteristic

Analysis Duration Analysis 100-car

Naturalistic Driving Data

Conclusion, Recommendation and Policy Implication

Chapter 2

Chapter 5 & 6

Chapter 7

Modeling

Introduction and Problem Statement _{Chapter 1}

Chapter 4

CHAPTER 2

E

SSENCE OF

D

RIVING

S

AFETY

A

NALYSES

The ultimate goal of an attention allocation analysis is to improve driving safety. Its connection within the crash occurrence can help identify the way that drivers interact with driving tasks, and probably the reason causing crashes. To elucidate the role of attention allocation in driving safety and crash prevention, the essence of crash analysis must be clarified. In this chapter, a comprehensive framework of safety analyses is constructed as Figure 2-1.

In general, a crash prone scenario represents a combination of risky factors within the driving stage of an accident chain. While driving in such crash-proneness scenarios, crashes were more likely, but not necessarily, to occur. There is a clear gap between the risky scenarios obtained from accident chain analysis and the crash occurrence. From the perspective of attention allocation, these risky scenarios may represent a condition that the drivers cannot perceive and process information adequately. The incomplete information perception would lead to higher chances of unexpected events, which induce reduced reaction time for drivers to response. In other word, the attention allocation analysis can illustrate in-depth characteristics of crashes from a chain perspective and help explore the last stage of an accident chain, namely the pre-crash stage.

The mechanism of drivers directing attention and processing information is the core of an accident chain. Certain factors in the pre-driving stage, such as drivers’ physical or psychological conditions, may affect the process of attention allocation. It does not only determine the habitual behavior pattern that drivers usually held, but also affect each driver’s capability of processing information. Seeing the limited attention resource, perceiving safety irrelevant information would decrease the attention resource being invested on the critical area for critical information. Additionally, drivers may evaluate the attention demand differently owing to the difference of their individual traits. Misjudging the attention demand of focal points may cause drivers allocating attention inappropriately.

To better understand the characteristics of attention allocation and its role in accident chain, section 2.1 and 2.2 reviewed the factors related to crash occurrence, including the crash-prone environment and drivers. Then, from an attention allocation perspective, contribution of these risk factors to attention demand was discussed in section 2.3. Considering the widely adoption of information system in recent days, section 2.4 illustrate its possible effect on attention allocation.

P h y si ca l C h ar ac te ri st ic · F at ig u e · A lc o h o l an d D ru g u sa g e · V is io n f ie ld · R ea ct io n c ap ab il it y E x p er ie n ce · T ra in in g · V io la ti o n · C ra sh · E d u ca ti o n · D ri v in g e x p o su re E n v ir o n m en t · W ea th er · P av em en t · L ig h t · S ig h t d is ta n ce T ra ff ic · C o m p le x it y · V eh ic le s p ee d · O cc u p an cy · In te ra ct io n w it h o th er v eh ic le s D ri v in g S ce n a ri o A tt en ti o n D em an d o f M u lt ip le F o ca l P o in ts D ri v er C h a ra ct er is ti c P ri o r-to -d ri v in g D ri v in g C ra sh A tt en ti o n A ll o ca ti o n · S al ie n ce · E ff o rt · E x p ec ta n cy · V al u e P er ce iv e N ec es sa ry In fo rm at io n P er ce iv e U n n ec es sa ry In fo rm at io n M is s N ec es sa ry In fo rm at io n R o ad w ay · G eo m et ry d es ig n · S ig n al · R o ad t y p e · S p ee d l im it · L an e w id th In fo rm at io n S y st em · In -v eh ic le in fo rm at io n s y st em · R o ad s ig n · C el l p h o n e · C o n v er sa ti o n D ec is io n M ak in g a n d M an eu v er A ct iv at io n E v en t O cc u rr en ce M en ta l T as k L o ad R is k h o m eo st as is M en ta l C ap ab il it y D ri v in g B eh av io r · P er so n al it y · D ri v in g c o n fi d en ce · S af et y A tt it u d e S it u at io n al U n aw ar en es s Po ss ib le C ra sh S af e d ri v in g Fig u re 2 -1 C ompre he nsive fr amew o rk of c ra sh an al ysis

2.1 Crash Pattern Analysis

Extracting patterns of crashes would help researchers understand the causality of a crash, and also find the way to prevent it. Comprehensive knowledge of contributing factors can provide clues to discover and reveal the nature of crash occurrence. In previous research, three types of crash pattern analysis were conducted: black spot analysis, crash type analysis, and crash severity analysis.

A black spot is any location that more crashes are expected to occur than other similar locations (Elvik 2008). It is a site-oriented approach that aims to extract recurrent crashes. Considerable research has been carried out to establish connections between frequency of crashes and various local characteristics of roadway and traffic (Chang and Chen 2005, Oh et al. 2006, Abdel-Aty and Pande 2007, Caliendo et al. 2007). The most common variables used to predict frequency of crash are the traffic volume (such as annual average daily traffic), roadway geometry (such as sight distance, horizontal alignment, vertical alignment and curvature), and environmental condition (such as weather, pavement, and light condition). Constructing the prediction model for black spot analysis is helpful in designing road and evaluating of safety improvement program. For example, Oh et al. (2006) examined factors associated with railroad crossing crashes and found average daily train traffic volume and proximity of crossings to commercial areas positively affect the crash occurrences. Chang and Chen (2005) used crash frequency of freeway in Taiwan to construct a non-parametric prediction model. They found that precipitation and daily traffic volume were the key determinant of crash frequency.

Instead of extracting recurrent crashes, crash type analysis focused on the uniqueness of crashes. Different crash types imply different interactions with road environment and with other vehicles. Rear-end, for example, is one common type of vehicle-to-vehicle crash. From the perspective of roadway characteristic, more signal phases, wider traffic island, higher speed limit, and higher number of lane increased the risk of rear-end crash (Chin and Quddus 2003, Yan et al. 2005, Wang and Abdel-Aty 2006). Moreover, Kostyniuk and Eby (1998) suggested that maneuver undertaken by frontal vehicle determined the occurrences of rear-end crashes. Any unexpected or unobserved maneuver undertaken by front vehicle creates greater danger of conflicts. To prevent conflict with frontal vehicle or other obstacles, drivers should maintain attention on the frontal side. Misallocating attention and failing to scan road ahead, particularly in congested traffic flow where drivers must frequently stop and go, increased rear-end crashes or conflict with fixed object (Golob and Recker 2003, Neyens and Boyle 2007a).

Another type of crash analysis is the exploration of crash severity. Vehicle type is the most important factors to determine the severity of a crash (Chang and Wang 2006). Concept of compatibility is proposed to evaluate the level of protection of each type of vehicle in a vehicle-to-vehicle crash (Mizuno and Kajzer 1999). Crashes which involved vehicles with similar compatibility were less serious. Mizuno and Kajzer (1999) suggested that SUV and mini car, which are the largest and the smallest vehicle in their research, are the two least competitive types of vehicle. Albertsson and Falkmer (2005) also suggested that probability of resulting in fatality and serious injury is higher in heavy vehicles related crashes than passenger vehicle crashes. To prevent possible serious crashes, drivers may be more concerned about certain types of vehicles on road, for example, the heavy vehicles.

In this section, the extraction of crash pattern is briefly reviewed. The scenarios explained the conditions in which drivers have increased risks of being involved in crashes, and possibly the driving scenario where drivers would be more likely to misallocate their attention. However, an unanswered question remains, namely the reason that crashes occur under specific conditions. The reality is that for each crash under certain risky conditions, there are numerous drivers who drive under identical conditions without experiencing crashes. The question thus arises of why different individuals react differently to identical conditions, resulting in different outcomes.

2.2 Driver Behavior Analysis

In addition to the analysis of factors closest to crash occurrences, the remote factors took place in the prior-to-driving stage should be analyzed (Elvik 2003, Wong and Chung 2007b, Wong and Chung 2007a, Wong and Chung 2008a, Wong and Chung 2008b). Driver is the most critical element within the prior-to-driving stage of accident chain. Age and gender are two observable variables which have been widely discussed. Regarding the age effect on driving, senior drivers have been found suffering degradation in driving skills, physical and cognitive conditions (Bayam et al. 2005). Accidents related to senior drivers usually resulted in losing the capability of situational awareness. Meanwhile, young drivers are usually considered as risky population and have the highest accident rate among all population (Clarke et al. 1998, Clarke et al. 1999). Gender is another important factor which distinguishes the accident patterns. Research conducted by Chang and Yeh (2007) stated that male drivers usually got involved in accidents due to their risky behavior while female drivers usually suffered accidents due to insufficient experience and skill.

Driver’s reaction and maneuver against external environment can be seen as the critical stage before crash occurrence. Provided that drivers are able to conduct safe maneuvers while driving in a risky scenario, the crashes are still preventable. Therefore, identifying how drivers normally drive becomes an important issue in clarifying the nature of crash occurrence. Questionnaire investigation was seen as a convenient approach for analyzing driving behavior. The driving behavior questionnaire (DBQ) was originally developed by Reason et al. (1990). Questionnaires containing 50 aberrant behaviors were distributed to obtain the frequency of driver undertaking specific aberrant behavior. After the factor analysis, three constructs of aberrant behaviors were extracted, which are harmless lapse, dangerous error, and violation. It was found that dangerous error decreased with the accumulation of exposure and experience. Parker et al. (2000) further divided the construct of violation into ordinary violation and aggressive violation. Senior driver were found conducting less aggressive violation but more lapses. It is suggested that senior population may face the degradation of mental capability which cause them unable to drive safety.

Among all human factors, psychological trait was one of the critical factors affecting risky driving behavior (Ulleberg and Rundmo 2003, Dahlen et al. 2005, Kim and Yamashita 2007). In order to discuss the decision making process of a driving behavior, Ulleberg and Rundmo (2003) adopted the Theory of Planned Behavior (TPB) and incorporated personality traits, attitudes towards safety and risk perception into Structure Equation Modeling (SEM) to discuss the risky driving behavior mechanism among young drivers. Based on this framework, Wong et al. (2010b) incorporated cost and benefit of conducting aberrant driving behavior. The results suggested that motorcyclists who have low riding confidence and traffic awareness deficiency usually over-focused on the object that pose threat and failed to observe surrounding traffic conditions. Based on the framework, Wong et al. (2010a) further examined the structural discrepancy that may exist in distinct groups of young motorcyclists by clustering the personality traits. Four types of young riders, risky, aggressive, conservative, and nervous, were extracted.

Clarifying the decision making process of conducting driving behavior help explain the accident chain. Combining the analyses of crash pattern driver behavior characteristics enables a deeper exploration of crashes. Yet, the real causalities were still not achieved. As mentioned in Wong et al. (2010b), aggressive motorcyclists tend to enjoy the utility of undertaking aberrant behavior. However, their experience and skill are able to adequately check surrounding traffic to prevent crashes from happening. In other words, a risky drivers driving in a crash-prone scenario were not

necessarily resulted in crashes. Obviously, missing link between accident chain and crash occurrence still exists. The critical issue in building up the link relies on the attention allocation while driving.

2.3 Attention Allocation

Attention allocation is the key to perceiving external information for making informed decisions to prevent crashes. Each risky driving scenario represents a set of information that drivers must gather from multiple information sources. In this section, a conceptual of driver attention allocation model is proposed based on the review of previous research.

2.3.1 Definition

The attention is a mental process of drivers’ mind and cognitive. Previous research defined the attention as “the process of concentrating or focusing limited cognitive resources to facilitate perception or mental activity” (Streff and Spradlin 2000, Regan et al. 2011). Key words of this definition are “concentration” and “limited cognitive resource”. That is, the attention must be a directive process, which allows cognitive resource to be invested on particular target.

Additionally, attention should be distinguished from the term of consciousness, although these two terms are highly related (Phaf et al. 1994, Treisman 2004, Koch and Tsuchiya 2006). There are four types of situations related to the consciousness and attention. One is attention without consciousness which represents the failure of perception; the case of “looked but failed to see”. It contains the behavior of directing attention to selected focus. Yet, attributes of the targeted object are neither identified nor perceived. Another is consciousness without attention. Objects outside the focal attention can be perceived without being selected as the focus through peripheral vision. In such condition, only partial attributes and information of the objects can be perceived. The third type is the maneuvering with no consciousness and no attention. This situation is usually caused by boredom or fatigue. The last type is the attention with consciousness, which is the one considered in this research. Thus, attention is defined as consciousness with focalization and concentration toward stimuli. In other words, once the attention is allocated, the information perception will be completely effective. Thus, based on the definition and the research scope, this study uses the “vision transition” as the proximity of attention allocation. Once a driver put his/her eyes one a specific target, the attention is allocated and invested.

In addition to the concentration of cognitive resource, another key term for defining attention is the resources being limited. Facing multiple sources of information, attention must be divided and allocated. The divided attention model proposed by Kahnemen (1973) stated that several activities can be focused on and carried out at the same time provided that their total effort is below the limit of available capacity. The capability of dividing attention resource to multiple targets is critical to situational awareness (Laberge et al. 2006, Creaser et al. 2007, de Waard et al. 2009, Marmeleira et al. 2009). To explain the divided attention concept, four principles of attention are mentioned. First, attention capacity is limited and varies from time to time. Available mental resources vary with the arousal level based on the physiology characteristics. Second, the amount of attention or mental resources allocated is based on the demand level of current activities. The more demanding an activity is, the more attention would be allocated to it. Third, attention is dividable. Fourth, attention is selective and controllable. A central policy exists for allocating attention to selected objects or activities. The framework of the divided attention model is illustrated in Figure 2-2.

Arousal Available Capacity Allocation Policy Possible Activities Response Evaluation of Demand on Capacity Enduring Dispositions Momentary Intentions Source: Kahnemen (1973) Figure 2-2 Model of divided attention

Four major elements are used to determine attention allocation policy in the model of divided attention: arousal, enduring dispositions, momentary intentions, and evaluation of demand on capacity. Arousal refers to factors such as physical condition, fatigue, or nervous tension that may activate the maximum attention capacity. An adequate level of arousal must be maintained. Under-arousal causes low attention capacity, whereas over-arousal impairs the ability to discriminate relevant objects from irrelevant objects. Enduring dispositions and momentary intentions reflect the characteristics of the external environment and behavioral intentions. Enduring dispositions represent state changes in the environment, such as deceleration of the

vehicle ahead, and reflect involuntary attention. Momentary intentions, in contrast, represent the intended attention allocation at that instant, such as searching for information using an in-vehicle information system. Finally, the feedback of attention allocation would continue to evaluate and adjust the arousal level and revise the allocation policy to fit the current situation.

2.3.2 Conceptual framework

Based on the concept of divided attention, Figure 2-3 illustrates the process of drivers dividing and allocating attention resources to different focal points for gathering information. The process comprises four stages, which are 1) accessing to the short-term memory, 2) allocating attention to focal points, 3) perceiving information, and 4) activating actions and updating memory.

Direct Vision to Specific Focal Point Attention Allocation Policy

Perceive Information from Chosen Focal Point Initial State of Short-term Memory

Process Information from Short-term Memory

Targeted Area (Full Information) Adjacent Area (Partial Information) Seeable Area (Low Information) Peripheral Vision

Perceive, Identify and Process Perceive and Identify

Targeted and Adjacent Area

· Existence of threats

· Maneuver around threats

· Contents of information

· Characteristics of orientation reaction

Seeable Area

· Existence of threats

· Characteristics of

orientation reaction

Information Perception

Information Is Enough for Achieving Intention?

Driving Maneuver Activation No Yes Update Memory Top-down Attention · Expectancy · Value Bottom-up Attention · Salience · Effort Driving Tasks

Figure 2-3 Process of driving attention allocation

In the first stage, short-term memory enables the maintenance of few information which is relevant with the ongoing tasks or activities. These information may be updated through the sustained attention and retrieved for making decisions or actions efficiently (Courtney 2010). From the perspective of driving, the short-term memory allows drivers to hold their comprehension of driving environment or other related statuses. It is also an important input for the attention allocation policy. Based on the

disposition of traffic flow, roadway and driving tasks that retrieved from short-term memory, drivers are able to evaluate the attention demand of each focal point and to direct visions to their intended area.

The second and the third stage determine the focal point chosen and the information perceived. Each driver has an attention allocation policy for determining the area to be glanced. For the chosen focal point, the targeted area enables the full information identification and perception. Drivers are able to identify the threat, monitor the movement, and predict its future status. In this level, the attention can be determined as the one with consciousness. Outside the central visual cone, drivers can still perceive partial information via peripheral vision. The amount of information perception degraded with the distance to central visual cone. To the information which is on the edge of peripheral vision, which is the seeable area, are barely comprehended. Drivers can only perceive the existence of an object with little information, such as the parking vehicle.

After the information perception, the last stage of attention allocation is the maneuvering and updating short-term memory. If drivers consider the necessary information is satisfyingly perceived and the current situation of traffic and other statuses are comprehended, the actions may be activated. Otherwise, another term of attention allocation should be undertaken for continuing the information perception. Noting that drivers may activate action without necessary information being completely perceived, such an uninformed action can lead to higher chances of unexpected events and possibly crashes. That is, misallocating attention or being inattention to critical information can result in dangerous situations.

The policy of attention allocation seems to be the key element of driving safety. It has been stated that the major distinction between experienced drivers from novices is the capability of utilizing their attention and mental resource (Konstantopoulos et al. 2010). Experienced drivers were considered having better knowledge of driving tasks and skilled attention allocation policy (Underwood et al. 2002a, Underwood et al. 2002b, Martens and Fox 2007, Nabatilan 2007, Borowsky et al. 2010). By contrast, novice drivers, who had immature mental models and limited rules of attention allocation, usually failed to anticipate hidden latent hazards and tend to commit more driving errors owing to a failure in attention allocation (Martens and Fox 2007, Chan et al. 2010). Moreover, Underwood et al. (2002b) suggested that novice drivers had more difficulties in controlling their vehicles. Therefore, they tended to focus more often on technical tasks and stare at the frontal side, instead of shifting vision around vehicles (Underwood et al. 2002b, Underwood et al. 2003a, Underwood 2007, Konstantopoulos et al. 2010).

All the works that have been done previously suggested the essential role of attention allocation policy. An adequate policy of attention allocation is necessary for safe driving (Shinar 2008). Exploring the way of drivers maintaining situational awareness through visual attention can be fruitful and potentially applicable for analyzing drivers’ behavior and preventing crashes. Therefore, to analyze the attention allocation in a more proper way, representations of attention were reviewed in section 2.3.3. Then, as stated in divided attention theory, dispositions of environmental conditions and driving tasks are critical determinants affecting the demand of attention. Section 2.3.4 reviewed the contributing factors that were adopted for attention allocation analysis.

2.3.3 Representations

Driver attention is not a manifest variable that can be measured directly. Thus, developing an appropriate representation of attention allocation is challenging. Nevertheless, various representations have been provided to analyze several aspects of attention allocation. There are three types of representation that previous research adopted to analyze the characteristics of attention allocation from various aspects.

The first type of representation aims to analyze the characteristics of a single focal point or target. Some studies utilized the portion of time that drivers spend looking at particular objects or areas as the representation of attention to show the importance of the areas (Underwood et al. 2002b, Underwood et al. 2003b, Nabatilan 2007, Di Stasi et al. 2009, Levin et al. 2009, Borowsky et al. 2010, Konstantopoulos et al. 2010, Dukic and Broberg 2012). Drivers usually spent more time on the focal point where they considered as one with higher risk of crashes.

Additionally, analyzing the duration (Falkmer and Gregersen 2001, Underwood et al. 2002a, Underwood et al. 2002b, Martens and Fox 2007, Di Stasi et al. 2009, Borowsky et al. 2010, Chan et al. 2010, Konstantopoulos et al. 2010, Dukic and Broberg 2012) and transition frequency (Salvucci and Liu 2002, Underwood et al. 2002b, Underwood et al. 2003b, Martens and Fox 2007, Kiefer and Hankey 2008, Di Stasi et al. 2009, Borowsky et al. 2010, Chan et al. 2010, Konstantopoulos et al. 2010) provided clues for identifying drivers’ mental status against driving tasks. For example, when facing mentally demanding tasks, drivers would increase their sampling rate for processing information more efficiently due to psychological pressure. Consequently, they would help short duration and high transition frequency while shifting vision among focal points (Chapman et al. 2002, Underwood et al. 2002a). On the other hand, provided the scenario is worsen, drivers would hold long

glance on the critical focal point with less vision transition and pay close attention to it (Underwood 2007).

Because presenting the process of drivers transiting visual fields among various focal points is not practical by employing the single-point approach, the second type of attention allocation representation adopted scan path approach to represent process of transiting vision among focal points (Underwood et al. 2003b). The scan path method examines multiple and sequential focal points to which drivers divert their glance. This method explores the detailed behavior of drivers shifting attention from one focal point to another. By extracting the scan path, it provides additional information on drivers’ sequential processes of attention allocation for maintaining situational awareness. The most common type of path is the one shifting vision toward the frontal side. Seeing that the drivers cannot perceive the status changes ahead while they glancing elsewhere, such a type of forward related path showed that the unawareness of leading traffic may urge drivers to transit vision back to the front (Brown et al. 2000).

In addition to analyzing the aggregated characteristics of attention allocation, from single point or scan path perspective, the third type of attention allocation representation is to calculate the probability of choosing specific focal point microscopically. Analyzing attention from a microscopic perspective can be seen as a mean to identify drivers’ dynamic behavior, particularly the attention allocation behavior, against the real world traffic (Laureshyn et al. 2010). In order to explains visual attention allocation in general for analyzing the situational awareness in dynamic environments, Wickens and his colleagues proposed a SEEV model for calculating the probability of choosing specific focal point under varying environment or task conditions. The model was originally design for a pilot’s attention allocation (Wickens et al. 2001, Wickens et al. 2003, Wickens and Thomas 2003, Miller et al. 2004). In recent years, this concept was adopted in the driving field (Horrey et al. 2006, Horrey and Wickens 2007, Werneke and Vollrath 2012). There are four constructs of attention demand included in the SEEV model, which are Salience, Effort, Expectancy and Value. Among the four constructs, only Effort provided negative effect on attention demand.

2.3.4 Contributing factors

Focal points or targets on roads do not attract drivers randomly. Instead, drivers, particularly experienced ones, usually direct their visions to focal points based the cues from environment or driving tasks (Falkmer and Gregersen 2001, Stanton and

Salmon 2009, Borowsky et al. 2010). Figure 2-4 shows the concept and definition of the four constructs used in the SEEV model. Among the four constructs, salience and effort are the two constructs representing bottom-up attributes of attention allocation. These attributes are exogenous and related to the characteristics of objects being targeted. Meanwhile, expectancy and value are the two constructs representing top-down attributes of attention allocation. These are endogenous attributes that characterize drivers’ knowledge-based skills.

Expectancy Value

Salience Effort

· Attention allocation sources

of higher task-relevant information bandwidth (event rate)

· Attention allocation to

sources of more value information to the task (higher and lower value)

· Attention allocation to events

or information with attention capture, e.g. flashing lights, bright lights, sudden onset, relative contrast, sound etc.

· Attention allocation between

two information sources with the shortest physical distance or the shortest time to assess the information

Visual Attention Allocation

Top-down

Bottom-up

Source: Werneke and Vollrath (2012); Wickens et al. (2001) Figure 2-4 The four components of the SEEV model

Expectancy was the first construct proposed in the SEEV model. It, as a top-down construct of attention demand attribute, determined the expected frequency of information appearing. Treating information perception as queuing behavior, Senders assumed that visual attention allocation was driven by the bandwidth of the information, which can be represented by the expected rate of status changes (Senders 1964, 1967, Moray 1986). Drivers would glance at the focal point where more stimuli, information or other status changes would appear (Verwey 2000, Blanco et al. 2006, Kiefer and Hankey 2008, Vashitz et al. 2008, Gershon et al. 2012). Moreover, traffic density and driving speed affect the level of interaction with other vehicles. When driving in heavy traffic (Werneke and Vollrath 2012) or high speed (Konstantopoulos and Crundall 2008), drivers would pay more attention to the frontal side for compensating the frequent status changes and short reaction time.

Based on the Senders’ model, Carbonell incorporated Value as another top-down construct for representing the importance and relevance of the information (Carbonell 1966, Carbonell et al. 1986). In driving tasks, the value of information is determined by their maneuvering status. Generally, drivers tend to look in the direction of future vehicle trajectories, i.e., where they expect the greatest number of threats to occur

(Martens and Fox 2007, Dukic and Broberg 2012, Gershon et al. 2012, Koustanaï et al. 2012, Lehtonen et al. 2012). For instance, moving forward constitutes a major driving activity. Hence, the frontal area attracts the most attention in almost all driving conditions (Underwood et al. 2003b, Nabatilan 2007, Underwood 2007, Konstantopoulos and Crundall 2008, Shinar 2008, Levin et al. 2009, Dukic and Broberg 2012). Changing lanes requires heightened attention to be invested in the adjacent lane (Salvucci and Liu 2002, Underwood et al. 2002b, Kiefer and Hankey 2008, Konstantopoulos and Crundall 2008). Entering an intersection compels drivers to look to both sides of the intersected roads (Summala et al. 1996, Konstantopoulos et al. 2010, Werneke and Vollrath 2012, 2013). In addition to the attention required for specific intended maneuvers, drivers allocate attention to surrounding areas to maintain awareness of traffic conditions and to prevent possible conflicts caused by other vehicles (Crundall et al. 2006).

In addition to maximizing the benefit (Value) of information perceived, drivers would also try to minimize the cost while gathering information, i.e., the Effort invested on particular targets (Kvålseth et al. 1976). This bottom-up construct determines the visual angle difference or the distance between two focal points (Wickens et al. 2003, Wickens and Thomas 2003, Horrey et al. 2006). Such a construct is important for representing the process of vision transition. Drivers, in general, tended to transit vision to focal point that is close to the current one. Underwood et al. (2003b) suggested that experienced drivers barely transited vision from one side of vehicle directly to another. By contrast, novice drivers sometimes undertake the vision transition across vehicles. In addition to the visual angles difference, drivers usually transit vision on a horizontal band (Falkmer and Gregersen 2001, Crundall et al. 2006, McIntyre et al. 2012). Shifting attention vertically required higher effort.

Another bottom-up construct of attention demand is the Salience, which represents the easiness of target being differentiated and identified from the background information. Drivers were easier to identify the target with higher contrast in color or conspicuity comparing with background (Koustanaï et al. 2012). For example, drivers were easier to identify motorcyclists wearing dark outfit in day time or bright outfit at night (Gershon et al. 2012). McIntyre and his colleague suggested that different color between brake lamp and rear lamp enable quicker reaction for the drivers on the rear side (McIntyre 2008, McIntyre et al. 2012). Billboard with high contrast and brightness attracted drivers’ attention more easily, and sometimes would cause dangers owing to distractions (Dukic et al. 2013). In addition to the conspicuity perspective of salience, a threat with unusual behavior and exterior can be viewed as a