Behavioural studies and safety evaluation based on video data

different types poses a question of how these data can be fused (Wender &

Dietmayer, 2007).

6.2. Behavioural studies and safety evaluation based on video data

The wider use of indirect traffic safety indicators has also been limited by the problems of establishing their validity, which requires collection of large datasets of both indirect indicators and accidents. For example, even though they are much more frequent than accidents, serious traffic conflicts are still too rare to allow collection of large samples using human observers. Although it is known (Hydén, 1987) that the relation between serious conflicts and injury and fatal accidents depends on the type of the conflicts (e.g. the angle of approach and types of road users involved), it is not always possible to split the available conflict data into many sub-categories as the number of conflicts (and accidents) in each group will be too few. A possible way to extend the accident dataset is to include slight injury and property-damage-only accidents that normally are not reported to the official statistics. Most probable, these accidents also correlate with the serious conflicts, but this hypothesis has never been possible to test as no data about such accidents was available. From this perspective, video analysis has a great potential to contribute to validation of the various safety indicators, since observations carried out over long periods will yield a large sample of the indicator measurements and the actual accidents, including the minor accidents.

The entire process of accident development can be studied and compared to the processes of the near-accidents and normal encounters. However, to make such studies, the observation periods have to be significantly extended (order of months and years compared to the current periods of a few days).

The construction of severity hierarchies requires the definition of some universal measure of severity. This is not a simple matter since this measure has to reflect both the accident risk and severity of possible consequences, make the produced hierarchy shape as close as possible to the “true” hierarchy shape and be operational enough to be applied to all the possible encounter types (with and without a collision or crossing course, different types of road users involved, etc.). There are some indications that human observers intuitively use some subjective measure of the severity that might fit these requirements. One of the conclusions from the international calibration study of traffic-conflict techniques from different countries (Grayson, 1984) is that “even the observers were instructed to use specific cues such as TTC or PET, they will incorporate other aspects of the situation as well. Although severity scaling is linked to objective measures, it also includes a subjective dimension. This results in common understanding of conflict severity, at least for trained observers”. Svensson, 1992, reports that the serious conflicts classified subjectively by observers correlate better with accidents than the conflicts defined strictly by the definition of the Swedish Traffic Conflicts Technique. The challenge, however, is to find an objective and operational measure that corresponds to the subjective severity judgements and validate it.

The advantage of using continuous indicators is that they allow us to study the development of an encounter as a process and test another approach to classifying the severity of encounters. Instead of using indicator values at a certain moment (as in the case of TTC_min, TA or PET), the entire indicator profiles, i.e., TTC or TAdv curves, can be examined. Such examination might reveal what typical shapes characterise

“normal” and “critical” situations, and the “nearness” of a profile to one or another

The wider use of indirect traffic safety indicators has also been limited by the problems of establishing their validity, which requires collection of large datasets of both indirect indicators and accidents. For example, even though they are much more frequent than accidents, serious traffic conflicts are still too rare to allow collection of large samples using human observers. Although it is known (Hydén, 1987) that the relation between serious conflicts and injury and fatal accidents depends on the type of the conflicts (e.g. the angle of approach and types of road users involved), it is not always possible to split the available conflict data into many sub-categories as the number of conflicts (and accidents) in each group will be too few. A possible way to extend the accident dataset is to include slight injury and property-damage-only accidents that normally are not reported to the official statistics. Most probable, these accidents also correlate with the serious conflicts, but this hypothesis has never been possible to test as no data about such accidents was available. From this perspective, video analysis has a great potential to contribute to validation of the various safety indicators, since observations carried out over long periods will yield a large sample of the indicator measurements and the actual accidents, including the minor accidents.

The entire process of accident development can be studied and compared to the processes of the near-accidents and normal encounters. However, to make such studies, the observation periods have to be significantly extended (order of months and years compared to the current periods of a few days).

The construction of severity hierarchies requires the definition of some universal measure of severity. This is not a simple matter since this measure has to reflect both the accident risk and severity of possible consequences, make the produced hierarchy shape as close as possible to the “true” hierarchy shape and be operational enough to be applied to all the possible encounter types (with and without a collision or crossing course, different types of road users involved, etc.). There are some indications that human observers intuitively use some subjective measure of the severity that might fit these requirements. One of the conclusions from the international calibration study of traffic-conflict techniques from different countries (Grayson, 1984) is that “even the observers were instructed to use specific cues such as TTC or PET, they will incorporate other aspects of the situation as well. Although severity scaling is linked to objective measures, it also includes a subjective dimension. This results in common understanding of conflict severity, at least for trained observers”. Svensson, 1992, reports that the serious conflicts classified subjectively by observers correlate better with accidents than the conflicts defined strictly by the definition of the Swedish Traffic Conflicts Technique. The challenge, however, is to find an objective and operational measure that corresponds to the subjective severity judgements and validate it.

The advantage of using continuous indicators is that they allow us to study the development of an encounter as a process and test another approach to classifying the severity of encounters. Instead of using indicator values at a certain moment (as in the case of TTC_min, TA or PET), the entire indicator profiles, i.e., TTC or TAdv curves, can be examined. Such examination might reveal what typical shapes characterise

“normal” and “critical” situations, and the “nearness” of a profile to one or another

“typical” shape may be used as a measure of the severity. Pattern recognition techniques could be a valuable tool in such analysis.

Some of the indicators that are important for estimation of the encounter severity, especially the severity of consequences in case of a collision, are difficult to collect using video analysis. This refers to the use of helmets by cyclists, motorcycle and moped drivers, use of safety belts by car drivers and passengers, age of road users, especially the vulnerable ones (pedestrians and cyclists), etc. In some cases, the necessary information can be extracted if a human observer looks through situations that have been selected automatically using some other criteria.

The proposed set of indicators describing an encounter is a first step towards finding a universal severity measure. Some work on validation of TTC and PET as severity measures in conflict situations has already been done (van der Horst, 1990, Hydén, 1987, Grayson, 1984). TTC was found to better reflect the severity than other indicators, for example, PET. It is still possible that TAdv combined with T₂ might be a better measure of severity than PET on its own. Other indicators, for example, TET, TIT or time-inversed TTC (1/TTC, Kiefer et al., 2005) should also be tested on a larger scale.

The definition of many of the proposed safety indicators is based on a concept of “the same paths and speed”, i.e., the motion of a road user has to be projected in the future. A human observer can (in most cases) make such a projection relatively easily, but it is difficult to explain exactly how this is done. It is possible that the price for seeming “easiness” is groove simplifications done, most probably, quite unconsciously, e.g. “compression” of a conflict zone into some vague “conflict point”, treatment of all approaching angles as if they were right angles, etc. When the projection is to be done automatically, clear and unambiguous algorithms are required. A simple assumption of travelling along a straight line does not work in the case of a road user making a turning manoeuvre, since in this case the potential collision point takes quite an unrealistic position. A possible approximation is to assume that road users actually follow the planned path, i.e., to use the known trajectory (if the indicator is calculated after the trajectories have been extracted). This may be misleading if the road user avoids a conflict by changing paths, for example taking a larger radius in a turn or changing lanes. Another alternative is to use an

“average” path, calculated from the trajectories of many road users making the same manoeuvre. The problem, however, is that in critical situations the paths might not follow the average pattern. A detailed analysis of critical situations might reveal when the deviation from the “average” pattern starts to develop during an encounter, and if the high severity of the situation can be detected before that moment, i.e., when the

“average” assumptions are still valid.

Different variations of TTC definitions were tested by van der Horst, 1990, for example based on assumptions of constant angular velocity and constant acceleration of a vehicle (this is supposed to represent a situation when a driver is no longer controlling the vehicle, and the steering wheel and the gas pedal positions are kept unchanged). The paths calculated with constant angular velocity easily take very peculiar shapes and lead outside the road. As for constant accelerations, the TTC

“typical” shape may be used as a measure of the severity. Pattern recognition techniques could be a valuable tool in such analysis.

Some of the indicators that are important for estimation of the encounter severity, especially the severity of consequences in case of a collision, are difficult to collect using video analysis. This refers to the use of helmets by cyclists, motorcycle and moped drivers, use of safety belts by car drivers and passengers, age of road users, especially the vulnerable ones (pedestrians and cyclists), etc. In some cases, the necessary information can be extracted if a human observer looks through situations that have been selected automatically using some other criteria.

The proposed set of indicators describing an encounter is a first step towards finding a universal severity measure. Some work on validation of TTC and PET as severity measures in conflict situations has already been done (van der Horst, 1990, Hydén, 1987, Grayson, 1984). TTC was found to better reflect the severity than other indicators, for example, PET. It is still possible that TAdv combined with T₂ might be a better measure of severity than PET on its own. Other indicators, for example, TET, TIT or time-inversed TTC (1/TTC, Kiefer et al., 2005) should also be tested on a larger scale.

The definition of many of the proposed safety indicators is based on a concept of “the same paths and speed”, i.e., the motion of a road user has to be projected in the future. A human observer can (in most cases) make such a projection relatively easily, but it is difficult to explain exactly how this is done. It is possible that the price for seeming “easiness” is groove simplifications done, most probably, quite unconsciously, e.g. “compression” of a conflict zone into some vague “conflict point”, treatment of all approaching angles as if they were right angles, etc. When the projection is to be done automatically, clear and unambiguous algorithms are required. A simple assumption of travelling along a straight line does not work in the case of a road user making a turning manoeuvre, since in this case the potential collision point takes quite an unrealistic position. A possible approximation is to assume that road users actually follow the planned path, i.e., to use the known trajectory (if the indicator is calculated after the trajectories have been extracted). This may be misleading if the road user avoids a conflict by changing paths, for example taking a larger radius in a turn or changing lanes. Another alternative is to use an

“average” path, calculated from the trajectories of many road users making the same manoeuvre. The problem, however, is that in critical situations the paths might not follow the average pattern. A detailed analysis of critical situations might reveal when the deviation from the “average” pattern starts to develop during an encounter, and if the high severity of the situation can be detected before that moment, i.e., when the

“average” assumptions are still valid.

Different variations of TTC definitions were tested by van der Horst, 1990, for example based on assumptions of constant angular velocity and constant acceleration of a vehicle (this is supposed to represent a situation when a driver is no longer controlling the vehicle, and the steering wheel and the gas pedal positions are kept unchanged). The paths calculated with constant angular velocity easily take very peculiar shapes and lead outside the road. As for constant accelerations, the TTC

values are still reasonable, but there is no clear evidence that the predictive power of TTC has improved. Still, further tests on the use of acceleration in calculation of the proposed indicator set are necessary.

Another theoretical problem is the assumption that an elementary event that can result in an accident is an encounter between two road users. This totally excludes situations with only one road user involved, even though single accidents are common and, for example, contribute 32 % of all the traffic fatalities in Sweden (SIKA, 2009).

The main difference in single accidents is that the factors contributing to the accident risk (e.g. fatigue causing a driver to fall asleep) cannot be attributed to some particular units of a road infrastructure (e.g. an intersection), but are spread over the entire network. Thus, it is not possible to study such accidents by making observations at a certain site; it has to be done, for example, from inside a vehicle. Some attempts to register traffic conflicts from vehicles are reported in the literature (e.g. Nygård, 1999, Risser, 1985). It is reasonable to assume that, similar to encounters, single-road-user-events belong to some kind of severity hierarchy, i.e., accidents, near-accidents when the road user manages to regain control of the vehicle at the very last moment and avoids the collision, and so on. It is possible to modify some of the indicators developed for encounter description so that they are applicable to single-road-user-events (e.g. the Time-to-Lane Crossing, TLC, is an extension of the TTC-concept and describes the time remaining for a road user to reach one of the lane boundaries – van Winsum et al., 2000). Still, further research is necessary on how these events may be integrated into the hierarchy based on encounters.

When a severity hierarchy is created, an important question is how it is to be interpreted. It is important to elaborate on what the whole shape and frequency of events at different levels represent. The severity hierarchies proposed earlier (Svensson

& Hydén, 2006, Svensson, 1998) include only events with a collision course. It is argued that interactions at fairly high severities may be positive from a safety point of view because they are frequent and severe enough to increase awareness. My suggestion is to include encounters without a collision course in the severity hierarchy, too, since such encounters also have the potential to become accidents if the spatial and temporal relation between the road users changes. It will be interesting to analyse whether these extended hierarchies can also be interpreted in a similar manner to a “collision course-only” hierarchy. With information about the encounter processes and the severity of these processes, it will be possible to formulate and test hypotheses on the interrelationships of design of the traffic environment, behaviour and risk. The final goal will be to have an operational and usable tool for safety estimation, similar to today’s traffic-conflict techniques, but with much higher degrees of validity, reliability and automation, which can be disseminated and used by traffic safety engineers on a daily basis.

The prospect of being able to collect data on road users over longer time periods and, possibly, over larger areas allows us to make new types of studies and observe completely new phenomena. Applied to vehicle traffic, it will be interesting to study the behaviour over longer road sections, for example adaptation of speed to the road geometry and elements of the infrastructure, interactions during lane change manoeuvres and near lane merging locations, etc. At the moment, our department is

values are still reasonable, but there is no clear evidence that the predictive power of TTC has improved. Still, further tests on the use of acceleration in calculation of the proposed indicator set are necessary.

Another theoretical problem is the assumption that an elementary event that can result in an accident is an encounter between two road users. This totally excludes situations with only one road user involved, even though single accidents are common and, for example, contribute 32 % of all the traffic fatalities in Sweden (SIKA, 2009).

The main difference in single accidents is that the factors contributing to the accident risk (e.g. fatigue causing a driver to fall asleep) cannot be attributed to some particular units of a road infrastructure (e.g. an intersection), but are spread over the entire network. Thus, it is not possible to study such accidents by making observations at a certain site; it has to be done, for example, from inside a vehicle. Some attempts to register traffic conflicts from vehicles are reported in the literature (e.g. Nygård, 1999, Risser, 1985). It is reasonable to assume that, similar to encounters, single-road-user-events belong to some kind of severity hierarchy, i.e., accidents, near-accidents when the road user manages to regain control of the vehicle at the very last moment and avoids the collision, and so on. It is possible to modify some of the indicators developed for encounter description so that they are applicable to single-road-user-events (e.g. the Time-to-Lane Crossing, TLC, is an extension of the TTC-concept and describes the time remaining for a road user to reach one of the lane boundaries – van Winsum et al., 2000). Still, further research is necessary on how these events may be integrated into the hierarchy based on encounters.

When a severity hierarchy is created, an important question is how it is to be interpreted. It is important to elaborate on what the whole shape and frequency of events at different levels represent. The severity hierarchies proposed earlier (Svensson

& Hydén, 2006, Svensson, 1998) include only events with a collision course. It is argued that interactions at fairly high severities may be positive from a safety point of view because they are frequent and severe enough to increase awareness. My suggestion is to include encounters without a collision course in the severity hierarchy, too, since such encounters also have the potential to become accidents if the spatial and temporal relation between the road users changes. It will be interesting to analyse whether these extended hierarchies can also be interpreted in a similar manner to a “collision course-only” hierarchy. With information about the encounter processes and the severity of these processes, it will be possible to formulate and test hypotheses on the interrelationships of design of the traffic environment, behaviour and risk. The final goal will be to have an operational and usable tool for safety estimation, similar to today’s traffic-conflict techniques, but with much higher degrees of validity, reliability and automation, which can be disseminated and used by traffic safety engineers on a daily basis.

The prospect of being able to collect data on road users over longer time periods and, possibly, over larger areas allows us to make new types of studies and observe completely new phenomena. Applied to vehicle traffic, it will be interesting to study the behaviour over longer road sections, for example adaptation of speed to the road geometry and elements of the infrastructure, interactions during lane change manoeuvres and near lane merging locations, etc. At the moment, our department is

launching several projects within the framework HASTA (Sustainable Attractive City – HASTA, 2009), aimed at studying the “overall transport quality” of a city where aspects like safety, health, security, accessibility, comfort, equality, participation and environment are treated integrally. The integral approach requires introduction of conceptually new indicators that still have to be developed. However, even now there is no doubt that many of the indicators will eventually be based on the micro-level behaviour of the “city users” and video analysis will be an indispensable tool for such data collection.

launching several projects within the framework HASTA (Sustainable Attractive City – HASTA, 2009), aimed at studying the “overall transport quality” of a city where aspects like safety, health, security, accessibility, comfort, equality, participation and environment are treated integrally. The integral approach requires introduction of conceptually new indicators that still have to be developed. However, even now there is no doubt that many of the indicators will eventually be based on the micro-level behaviour of the “city users” and video analysis will be an indispensable tool for such data collection.