Accident Analysis and Prevention 38 (2006) 379385

Estimating the severity of safety related behaviour

Ase Svensson , Christer Hyd n e
Department of Technology and Society, Lund Institute of Technology, Lund University, Box 118, SE-22100 Lund, Sweden Received 14 October 2005; accepted 18 October 2005

Abstract The aim of this work is to be a starting point for a more thorough description and analysis of safety related road user behaviour in order to better understand the different parts forming the trafc safety processes. The background is that it is problematic to use analysis of crash data and conict data in the everyday trafc safety work due to low occurrence rates and the focus on rather exceptional and unsuccessful events. A new framework must consider the following aspects: (1) The importance of feedback to the road users. (2) Inclusion of more frequent events, normal road user behaviours and the possibility to link them to a severity dimension. (3) Prediction of safety/unsafety based on the more frequent events. By constructing severity hierarchies based on a uniform severity dimension (Time to Accident/Conicting Speed value) it is possible to both describe the closeness to a crash and to get a comprehensive understanding of the connection between behaviour and safety by both considering unsuccessful and successful interactive situations. These severity hierarchies would make it possible to consider road users expectations due to feedback and estimate its safety relevance. 2005 Elsevier Ltd. All rights reserved.
Keywords: Trafc; Safety; Behaviour; Conicts; Interaction; Feedback

1. Introduction The aim of the work behind this paper is to extend the trafc safety assessment concept to also include normal road user behaviours, thus not only exceptional behaviours such as those leading to crashes and/or serious conicts. The goal is to provide a framework for a more thorough description and analysis of road user behaviour in order to better understand what we dene as the trafc safety processes, i.e. the interactional processes that dene events of different severity. 2. Background 2.1. Interaction Trafc is interactionall events in trafc contain some kind of interaction but of course to varying extent. There is interaction between road users and there is always interaction between the road user and the road environment. In this paper the term

interaction is restricted to the relation between road users. The interaction between road users can be described as a continuum of safety related events (see Fig. 1). This pyramid shows how few and exceptional those events are that we usually base our safety estimates on, i.e. the crashes, rarely also including the serious conicts. 2.2. Need for surrogate measures The traditional way of approaching trafc safety has mainly been concerned with the occurrence of trafc crashes and their consequences. There are, however, disadvantages with the use of crash data analyses and these have been discussed extensively in several reports, e.g. Englund et al. (1998), Grayson and Hakkert (1987). (1) Crashes are rare events and are therefore associated with the random variation inherent in small numbers. (2) Not all crashes are reported and the level of reporting is unevenly distributed with regard to, e.g. type of road users involved, location, severity of injuries, etc. (Berntman et al., 1995). (3) The behavioural or situational aspects of the events are not covered in police crash data (Berntman, 1994). Crashes are also exceptional in the sense that they are a collection of events where all alternatives to handle the situation safely, have vanished one by

Corresponding author. Tel.: +46 46 2229125; fax: +46 46 123272. E-mail addresses: Ase.Svensson@tft.lth.se (A. Svensson), Christer.Hyden@tft.lth.se (C. Hyd n). e 0001-4575/$ see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.aap.2005.10.009

380

A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e

strong evasive action to avoid a crash. Experience with the TCT has, nevertheless, shown that it is possible to include less severe events than crashes, i.e. serious conicts, and reach better understanding of the trafc safety process. 3. Extension of the concept As the task here is to try to explain the relationship between road user behaviour and safety this implies an unambiguous need for widening the scope of trafc safety and safety related events. There are at least three fundamental issues that are important to consider, to link and to interpret when structuring a new framework. The importance of feedback to the road users. Inclusion of more frequent events, normal road user behaviours and the possibility to link them to a severity dimension. Improve prediction of safety/unsafety. 3.1. The importance of feedback to the road users What is it that makes one trafc environment more crash prone than another? A basic hypothesis in this paper is that the feedback to the road users can be an important explanatory factor. The occurrence of crashes can be due to lack of feedback but also due to the fact that existing feedback is misleading or perhaps incorrectly interpreted. The importance of feedback to road users has not least been acknowledged when trafc education for children is discussed. According to Thomson et al. (1996) referred to by Whitebread and Neilson (1996) pedestrians require a range of fundamental skills to interact safely in trafc. The pedestrian has to, among many other things, make judgements of whether the crossing place is safe or not by co-ordinating past experience, present information and predictions about the future. Children lack this cognitive ability and they have due to obvious reasons not yet had the time to achieve feedback based on previous experiences in trafc. Thus, practical training in trafc is crucial. Crash statistics also clearly show the need for practical training and the importance of feedback. Young drivers with a fresh drivers license is a road user group with high crash risks. Experience obviously plays an important role here. It is, however, according to Evans (1991) reasonable to in addition assume that the over involvement by young road users in trafc crashes must involve more than lack of driving experience as the tendency is the same for pedestrians as for car drivers at that age. Feedback from interactions is most likely a very important part of the learning process and the more obvious feedback the road users have got, the more inuence it has on their behaviour in similar situations/environments. 3.2. Inclusion of normal road user behaviour 3.2.1. Criteria of events in the framework We are looking for a framework that handles predened events that are much more frequent than injury crashes and seri-

Fig. 1. The pyramidthe interaction between road users as a continuum of events (Hyd n, 1987). e

one. This is indeed exceptional compared to most other events that actually are handled safely (though to different degree). Thus, we need to get a more comprehensive understanding of the connection between behaviour and safety by both considering unsuccessful and successful interactive situations. The need for surrogates or complementary methods for crash data analysis is consequently highthe Trafc Conicts Technique (TCT) is such a method. A conict is a situation where two or more road users approach each other in time and space to such an extent that a collision is imminent if their movements remain unchanged. The development of TCT has shown that serious conicts contain most of the qualications lacking with crash data analysis. Serious conicts do for instance possess the quality of being an indicator of a breakdown in the interactiona breakdown that could correspond to the breakdown in the interaction preceding a crash. A serious conict is also, like the crash, a situation that nobody puts him/herself into deliberatelythe situation is perceived as being too threatening (Hyd n and St hl, 1979). e a The relationship between serious conicts and injury crashes reported by the police has been elaborated on and established through two validation studies. Hyd n (1987) deals with three e samples of data. The product validation part produces a set of conversion factors, i.e. establishes the relationship between the number of crashes and the number of serious conicts. In the process validation part analyses are conducted regarding similarities of the processes preceding crashes and serious conicts. Analyses showed big similarities between crashes and conicts when the comparison was based on Time to Accident (TA) values and Conicting Speed (CS). It also showed that the distributions of different types of evasive action were very equal for crashes and conicts. In Svensson (1992) analyses on the product validation of the Swedish TCT show that at lower crash frequencies it is preferable to use conicts instead of crashes when estimating the expected number of crashes. For further information about the Swedish TCT and other TCTs see also, e.g. Grayson (1984). Many of the shortcomings in crash data analyses are provided for with the use of TCT, but not all. Sometimes also the serious conicts are too few to obtain statistically signicant estimates at assessment studies. The analyses of serious conicts do also have the same angle of approach as the crash data analysis, i.e. the primary focus is set on rather exceptional and unsuccessful events; unsuccessful in the sense that road users have to take

 A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e

381

ous conicts; events that are not as exceptional as injury crashes and serious conicts but still have a logical link to crashes and serious conicts. The framework and the selected events must make it possible to make analyses of the feedback effect on an individual level as well as on an aggregated level, e.g. of locations, of type of manoeuvre, etc. To be able to make comparisons between different locations with, for instance, different geometrical design, the severity of the predened events and the distribution of these events must be described in a reliable and uniform way. This distribution of events, a severity hierarchy, is a feasible solution where the distribution of events over severity will provide us with information regarding the trafc safety situation at a location. Another prerequisite is that the events should describe behaviours that can be related to feedback, i.e. to road users expectations of the prevailing situation occurring in a specic trafc environment. Finally, the prediction of safety/unsafety should be improved by using these events. 3.2.2. Severity and severity hierarchy A collision presupposes a collision course. It therefore seems logical to dene the common denominator for events to be included in the severity hierarchy, as those events where the road users move on a collision course. The aim must, furthermore, be to construct a severity hierarchy for trafc events so that for each event a severity, i.e. the events closeness to a crash, can be estimated. One feasible, and here chosen, possibility is to describe an events severity by its Conicting Speed and Time to Accident value. The Conicting Speed is the speed of the road user taking evasive action, for whom the TA value is estimated, at the moment just before the start of the evasive action. The Time to Accident is the time that remains to a crash from the moment that one of the road users starts an evasive action if they had continued with unchanged speeds and directions. These measures have been applied in the Swedish TCT. For clarication it might be added that situations with low PET (PostEncroachment-Time) values are included in the Swedish TCT, however on the precondition that the involved road users behave as if they were moving on a collision course. It has been shown (Hyd n, 1987) that serious conicts, as dened in the Swedish e TCT, are events that represent a logical continuation of crashes on a severity scale. It is therefore reasonable to assume that, using the same severity denition, less severe events could logically follow on the serious conicts. Severity, described by an events TA/CS value, does not refer to the known outcome after the evasive action but to the severity of the event an innitesimal unit of time before the evasive action. At this very moment we can say that an unknown event with a certain location in the severity hierarchy has a certain closeness to an injury crash, i.e. a certain probability of resulting in an injury crash. The outcome in the form of a crash or not then depends on the success of the evasive action and of course on the local characteristics. In the work of identifying factors with regard to risk causation we would be much better off starting from the much more

Fig. 2. TA/CS graph dening the different severity levels. There is a continuation towards lower severity levels. Severity level 1(not shown) intersects the X-axis at TA = 13.0 s (Svensson, 1998).

frequent normal behaviours, than the rare and unique crashes. This, however, presupposes that all events with a TA/CS value are parts of the same trafc safety process as the crashes and serious conicts, i.e. that there is a link between normal road user behaviour and critical events with crashes at the end of the scale. To conclude: Relevant events in the trafc safety process are henceforth called interactions. Interactions are characterised by collision course and the severity is described by the TA/CS values. The severity distribution of the interactions is analysed by the construction of severity hierarchies (see Figs. 2 and 3). The understanding of different hierarchy shapes together with information about absolute numbers on different severity levels will increase the prospect of making reliable predictions about safe/unsafe environments and make it possible to consider road users expectations due to feedback.

Fig. 3. Severity hierarchy with severity levels corresponding to the ones dened in Fig. 2 (Svensson, 1998).

382

A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e

Fig. 4. Example of a possible hierarchy shape (Svensson, 1998). Fig. 5. Model for connecting the road user behaviour to predictions about safety and unsafety.

Hyd n (1987) indicated with his pyramid (Fig. 1) that there e would be a continuous increase in the number of events the further away you come from the most severe events. This is obviously true when we look at all events. If we, however, restrict ourselves to interactions characterised by collision course, then it is likely that the shape is different. As with accepted gaps, speeds, etc. it is most likely that one will nd a distribution that will be similar to a normal distribution, i.e. the severity hierarchy will be peaked on both sides (Fig. 4). The hypothesis is that different hierarchies will have different degree of accumulation of interactions allocated to different parts of the hierarchy. The shape and position of the convexity, i.e. the part of the hierarchy where most events are located, could possibly reect the most frequent interactive road user behaviour for that location/manoeuvre/road user type. The convexity could be interpreted as representing the road users optimisation of their desires to keep a high mobility standard, to maintain safety margins and to maintain comfort. 3.3. Prediction A new framework should improve the conditions of making reliable safety/unsafety predictions. If we by safety mean lack of crashes or serious conicts then we would have to equate unsafety with the presence of crashes and/or serious conicts. This denition is, however, not adequate for at least the two different reasons mentioned earlier, i.e. crashes are rare and they do not offer any understanding of the causality. A more thorough understanding could be achieved by applying a concept where it would be possible to identify the preconditions for a safe/unsafe environment. Such a concept prerequisites a broader angle of approach than merely stating whether crashes occur or not. 4. Model Fig. 5 below shows what our model for the relation between road user behaviour, shape of the severity hierarchy, feedback and predictions for safety and unsafety looks like.

4.1. Relations in the model 4.1.1. Interactions with TA/CS value severity hierarchy prediction The information we get from each interaction and its severity, is how close this particular interaction was to a crash, i.e. how imminent the danger was. Also the different parts of the hierarchy (clusters of interactions) contain information, like; the signicance of the predominant behaviour when it comes to describing the probability for the location to produce safe/unsafe interactive situations; or the signicance of the absolute numbers on the different severity levels. The uttermost interest, however, lies in the understanding of different hierarchy shapes, i.e. different compositions of interactions with different severity. If we were to understand the relations between the different levels in the hierarchy and their relevance for safety we would be in a much better position for exploring the pre-conditions for safe/unsafe road user behaviour and safe/unsafe locations, i.e. to make more reliable future predictions. 4.1.2. Interactions with TA/CS value severity hierarchy feedback The severity described by the TA/CS value can be interpreted as a reection of road users expectations of the situation/location. These expectations are formed by earlier experiences in similar situations/locations. Crashes and serious conicts are probably too rare to produce efcient feedback to an individual road user. How often will the road user be facing this very specic situation again? The feedback is obvious but not operational. For the individual road user it is important that there are more events from which he/she can learn. There is of course an interplay between feedback, expectations and predominant behaviour. Road users get feedback from being involved in different types of situations; road users expectations are then inuenced by this feedback; these expectations then set the scene for the composition of behaviours at a location; and so on. When the severity becomes too high it is primarily because the road user(s) are not prepared to interact due to expectations based on earlier feedback. A vehicle driver approaching a green

 A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e

383

Table 1 Interaction frequency per hour per severity level for interactions involving straight ahead driving vehicles versus pedestrians Severity level Interaction frequency per severity level for different category cells VSp 30 29 28 27 26 25 DSp

Fig. 6. Interaction frequency (interactions per observation hour) for different severity levels. Straight ahead driving vehicles versus pedestrians. The pedestrian is taking evasive action. A non-signalised intersection (DSp) and a signalised intersection (VSp). Please note that the severity level here is on the x-axis compared to on the y-axis in Figs. 3 and 4 (Svensson, 1998). The severity of an event increases with increased severity level.

2.28E-05 2.28E-05 0 0

0 0.94

signal at an intersection expects crossing pedestrians to stop for the red signal; that is what pedestrians usually do. A situation with a too high severity might occur at the rare occasion when a pedestrian actually crosses on red. G rder (1982) showed that a the proportion of jaywalkers at trafc signals increased with increased city size, while at the same time the higher the proportion of jaywalkers was the lower the risk of being involved in an injury crash was. The vehicle drivers expectations are violated and he/she is therefore not prepared to act. By analysing all interactions with a TA/CS value, from the lowest to the highest severities, at a location and by analysing the shape of the severity hierarchy, the distribution of the road users expectations the predominant behaviour can be discussed. By analysing the different shapes of severity hierarchies based on interactions at different types of locations, involving different types of manoeuvres, involving different types of road users, it would be possible to identify different groups of interactions that (1) give relevant feedback; (2) give no feedback or even give misleading feedback. Thus, it would be possible to dene locations and derived behaviours as safe/unsafe due to the quality of the feedback to the road users. 4.1.3. Feedback shape of the severity hierarchy prediction In this section we will discuss the shape of the hierarchy based on interactions collected at two different intersections, one signalised and one non-signalised intersection. The aim is (1) to try to understand what the different parts in the hierarchy represent; (2) to try to identify if there are preconditions for relevant feedback or misleading feedback to the road users; (3) to try and estimate if it would be feasible to make predictions based on the feedback at the two different locations. The interactions, in Fig. 6 and accompanying Table 1, involve straight ahead driving vehicles and pedestrians, where it is the pedestrian who is taking evasive action. The two different curves, VSp and DSp in Fig. 6, represent interactions at a signalised respective at a non-signalised intersection. Due to the fact that these analyses only are based on two specic intersections and for specic manoeuvres, the results are of course uncertain and we have to be very cautious when trying to make any kind of generalisation.

The pedestrian is taking evasive action. A non-signalised intersection (DSp) and a signalised intersection (VSp). Only the severity levels with the highest severities are included. The severity of an event increases with increased severity level. The interactions on severity levels 27 and 28 are injury crashes reported by the police.

For the highest severities, severity levels 2630, there are difculties to see any difference due to the low frequencies. This is instead shown in Table 1 below. There are indications of great differences in the shapes of the hierarchies, which obviously have to be due to differences in the road user behaviour at a signalised and at a non-signalised intersection. There is a difference regarding the frequency of events at different severity levels. The location and the extension of the convexity, i.e. the part of the hierarchy where most events are located, is also different for the two hierarchies. 4.1.4. Events of the highest severities With our chosen severity dimension the highest severity levels are level 26 and more severe (see Fig. 2). The events at these severity levels consist of serious conicts and injury crashes. They do give very relevant feedback to the involved road users. The question is, however, if they are frequent enough to produce efcient feedback. The indication is that the signalised intersection seems to produce more occasional events of the highest severities than the non-signalised intersection. 4.1.5. Events of fairly high severities The second group of interactions, those of fairly high severities, is located at severity levels 2025. Even these interactions are characterised by closeness in time and space thus still having a strong closeness to an unsafety dimension. At these levels the safety margin is so small that the situation easily can turn into a much more critical event. There are two possible interpretations of the fact that there are many interactions at these fairly high severities, either (a) something good because there is behavioural feedback regarding the necessity of increasing the awareness at interactions with fairly high severity, or (b) something bad because interactions at these high severities are always associated with an element of surprise and they are so close that they easily can turn into more serious events, i.e. crashes. Interactions at fairly high severities are very frequent; they constitute the predominant behaviour at the non-signalised intersection, while there are no crashes or serious conicts. At the signalised intersection on the other hand, those interactions with

384

A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e

fairly high severity do not seem to exist while crashes do. This supports the alternative interpretation that these interactions can be positive because they are frequent and severe enough to produce efcient feedback to the involved road users, but not severe enough to result in crashes. 4.1.6. Events of lower severities For the group of interactions located further down in the severity hierarchy the danger is not very imminent. We are now talking about events located at severity level 19 or less severe (see Fig. 2). The risk that one of these interactions results in a crash is very small, or in terms of TA/CS values; there is no closeness. Nevertheless, if the majority of events are located at these levels, it could have different interpretations and implications on safety. One interpretation could be that this is good for trafc safety as the normal road user behaviour is to divert from collision course at an early stage. From a feedback point of view, however, they can be considered as negative due to the lack of other feedback than everything went ne as usual. These situations are not perceived as interactions. Road users are reacting on the signal not on the presence of other road users on collision course. These situations, therefore, could place the road user in a difcult position once there is an interaction that requires a good preparedness for safe interaction, e.g. when somebody violates a red signal. A consequent hypothesis is therefore that a trafc environment where the majority of interactions consists of these types of interactions with poor feedback, can be an indicator of severe safety problems. 4.1.7. Convexity The form of the convexity, i.e. the part of the hierarchy where most events are located, can evidently range from being narrow with regard to the extension over severity levels, as for the nonsignalised intersection, to being more widely spread over several severity levels, the signalised intersection. The latter could be an indication of road users difculty to interpret and decide upon signs of possible threat. It can also be seen as a conrmation of the safety effect of signalisation; the priority rules are clear and the intention of the pedestrians is to stop but due to differences in the individual safety margins the evasive action is taken at different closeness to the signal (or rather to the interacting ow of motor vehicles). 4.1.8. The shape of the whole hierarchy The total shape of the severity hierarchy can be interpreted as the distribution of individual safety marginssafety margins that differ due to each individuals unique acceptance of comfortable margins in time and space at interactions, and due to time of detection. According to N at nen and Summalas (1976) discusa a sions about the zero-risk theory, these margins can also depend on other considerations than safety, such as the wish to maintain a certain speed or the wish to conserve energy and comfort. In addition to these motives behind the safety margins chosen, we have to consider the complex relation to feedback. The interactions at the signalised intersection (dotted line in Fig. 6) do at rst sight seem to be the result of safe behaviourmost interactions are handled in due time before

they become critical. These margins are on the other hand not the result of deliberate considerations but rather the result of interactions with the signal itself. We do also have the information that pedestrians do get killed and seriously injured at this type of intersectionthe narrow peak up to the highest severities (Table 1). There is a safety problem at the signalised intersection at the same time as most interactions are located towards the lower severities while there is a lack of interactions at the fairly high severities. The shape of the interactions at the nonsignalised intersection (solid line in Fig. 6) gives certainly the analyser the feeling of unsafetymost interactions are located at fairly high severities. However, there are no indicators of safety problems at this location even though most interactions take place at fairly high severities. A reasonable assumption about this latter group of interactions is that here the mobility and safety desires of the involved road users have been balanced. The analyses here point at the necessity of interpreting the different parts of the hierarchy shapes in order to better understand the preconditions for good and relevant feedback and their safety implications. However, when it comes to the prospect of making reliable safety predictions it is presumably an advantage if it can be based on the shape of the whole severity hierarchy. 5. A need for a behaviour-based framework in exploring trafc processes By continuing working according to the outlined approach in this paper we hope that the concept of analysing the shape of severity hierarchies in the near future can be used: in describing differences in road user behaviour; to improve the understanding of road user behaviour; for predicting the frequency of the most severe events from information about less severe events; to learn about the importance of feedback and different types of feedback; for formulating trafc safety strategies. References
Berntman, M., 1994. Metoder f r insamling av uppgifter om sv rt trako a skadade n gra k llor och tekniker. Bulletin 6. Department of Trafc a a Planning and Engineering. Road Construction, Lund University, Lund, Sweden. Berntman, M., Berntman, L., Nilsson, S-A., 1995. Traks kerhetsproblem i a Lunds kommun Sjukhusets och polisens skadestatistik 1988-10 199309. Bulletin 7. Department of Trafc Planning and Engineering. Road Construction, Lund University, Lund, Sweden. Englund, A., Gregersen, N.P., Hyd n, C., L vsund, P., Aberg, L., 1998. e o Traks kerhet En kunskaps versikt. KFB. Art. nr. 6319. Printed by a o Studentlitteratur, Lund, Sweden. Evans, L., 1991. Trafc Safety and the Driver. Library of Congress c Catalog, Van Nostrand Reinhold. Grayson, G.B. (Ed.), 1984. The Malm Study. A calibration of trafc conict o techniques. Institute for Road Safety Research SWOV, Leidschendam, the Netherlands. Grayson, G.B., Hakkert, A.S., 1987. In: Rothengatter, J.A., de Bruinj, R.A. (Eds.), Accident analysis and conict behaviour. Road Users and trafc safety. Van Gorkum Publishers, The Netherlands.

 A. Svensson, C. Hyd n / Accident Analysis and Prevention 38 (2006) 379385 e G rder, P., 1982. G endes s kerhet i traksignaler. Department of Trafc a a a Planning and Engineering. Lund University, Lund, Sweden. Hyd n, C., 1987. The development of a method for trafc safety evaluation: e The Swedish Trafc Conicts Technique. Department of Trafc Planning and Engineering. Lund University, Lund, Sweden. a Hyd n, C., St hl, A., 1979. Atg rder i trakmilj n f r att oka oskyddade e a o o trakanters s kerhet p stomgator. Department of Trafc Planning and a a Engineering. Lund University, Lund, Sweden. N at nen, R., Summala, H., 1976. Road User Behaviour and Trafc Accia a dents. North-Holland/American Elsevier, Amsterdam and New York. Svensson, A., 1992. Vidareutveckling och validering av den svenska konikttekniken. Department of Trafc Planning and Engineering. Lund University, Lund, Sweden.

385

Svensson, A., 1998. A method for analysing the trafc process in a safety perspective. Department of Trafc Planning and Engineering. Lund University, Lund, Sweden. Thomson, J.A., Tolmie, A., Foot, H.C., McLaren, B., 1996. Child development and the aims of road safety education a review and analysis. Road Safety Research Report No. 1, Dept. of Transport Norwich, HMSO. Whitebread, D., Neilson, K., 1996. Metacognitive and cognitive style elements of childrens pedestrian skills: some research ndings and implications for road safety training. In: Proceedings of the Conference Road Safety in Europe, Birmingham, UK, September 911, 1996. VTI konferens. No. 7A, Part 3.