Bhopal ; space exploration, e. Summerland leisure centre fire ; rail crashes, e. Ladbroke Grove ; and crowd crushes, e. Love Parade Whole systems are rarely evaluated as organ- isations search for technical causes, seeking to assign responsibility and develop pre- vention techniques at the expense of addressing the underlying root causes e. Canter et al. Email: r. Challenger and C.
Whilst we agree with this fundamental principle, we believe these theories are under-specified both theoretically and practically. Is it enough to argue simply that unanticipated events may come together to create disasters? Can we use design principles to identify what violations may contribute to systems failure? Is there a way of providing greater specification of past and future issues for theoretical and practical purposes? In this paper, we attempt to provide greater specificity by presenting a socio-tech- nical systems framework and underpinning design principles for analysing disasters.
Although our approach should be applicable across different domains, we focus on crowd-related disasters, following our previous work for the Cabinet Office Challenger et al.
The objectives of this paper are as follows:. To present our socio-technical systems framework and design principles for analys- ing disasters. To make theoretical contributions to the academic literatures on disasters and socio-technical thinking by providing greater specification of the kinds of factors that come together to create major problems. To suggest ways forward for the practice of those involved in organising, planning and managing crowd events. A socio-technical systems framework and design principles Socio-technical systems theory advocates when designing and operating any new system it is critical to focus on and optimise both technical and social factors e.
Crowd disasters: a socio-technical systems perspective Downloaded by [Rose Challenger] at 17 November Figure 1. A socio-technical systems perspective Challenger et al. It is inevitable that changes to one part of a system will necessi- tate subsequent changes to other parts; thereby, to optimise success, the system should be considered holistically e. Thus, people, processes and procedures, goals, culture, technology, and buildings and infrastructure should all be viewed as interdependent and given joint consider- ation, as illustrated in our socio-technical systems framework Figure 1.
Further- more, it is highly unlikely that any individual or group will understand all the component parts when considering the system overall. Therefore, new systems design should involve multiple stakeholders with a complementary range of knowl- edge and expertise, including end-users, managers, human resource experts, designers and clients e.
Clegg et al. End-user participation in, and owner- ship of, systems design and implementation is critical e. There are a number of interrelated principles for socio-technical systems design Cherns, , ; Clegg, that underpin our approach Table 1. They have four primary functions: to highlight issues requiring particular attention in the design process; to stress the need for a series of interrelated perspectives on design; to provide a potential framework for systems analysis; and to enable predictions about future systems operation Clegg, Method To assess the explanatory power of our socio-technical approach, we analysed retro- spectively a number of crowd-related disasters.
Our selection criteria were threefold. The disasters must: 1 involve crowds directly; 2 be independently and rigorously reviewed by formal public inquiries and peer-reviewed publications ; and 3 have similarities and differences e. Clegg Table 1. Principles of socio-technical systems design adapted from Clegg, Meta-principles capture an overall view of systems design 1 Design is systemic A system comprises a range of interrelated factors and should be designed to optimise social and technical concerns jointly 2 Values and mindsets are central to design Underlying values and mindsets strongly influence systems design and operation 3 Design involves making choices Design choices are interdependent and exist on many dimensions, e.
All three involved crowds directly and were subject to independent public inquiries leading respectively to the Taylor Report Taylor, , , the Fennell Report Fennell, , and the Popplewell Inquiry Popplewell, Whilst these official reports were used as the primary data source, relevant academic papers were also examined. Delayed journeys meant many of the 24, Liverpool supporters arrived late with only 30 minutes to enter the ground before the 3 p.
This led to a rush of over supporters into the Leppings Lane terrace directly behind the goal, via a steep, narrow tunnel Taylor, Several side pens remained half empty whilst the central two pens 3 and 4 , with the easiest immediate access, became severely overcrowded. Mass crushing occurred and a barrier A collapsed under immense crowd pressure Scraton, Some fans started to climb the perimeter fence to escape the crush, whilst others forced open a small gate in the fencing and escaped onto the pitch.
Two per- imeter gates were opened and fans evacuated onto the pitch. In total, 95 people died and over were injured Taylor, , However, he also acknowledged the influence of issues such as complacency, poor facilities and ground conditions, concerns over hooliganism, and poor leadership Taylor, Indeed, mapping the findings of our analysis onto our systems framework it becomes clear that problems occurred across the six interrelated factors Figure 2 , underpinned by the design principles, as detailed below.
Primarily, we believe the mindsets and values at Hillsborough were incompatible with systems thinking principle 2. The overall attitude was one of complacency. The Hillsborough football stadium disaster from a systems perspective. In line with this atti- tude, there was a fundamental failure to learn lessons from previous incidents prin- ciple Indeed, crushing had been reported at Hillsborough during the FA Cup semi-final, when overcrowding in pens 3 and 4 was so severe that police blocked off the tunnel leading to them.
They shared their stories in a sincere effort to help their colleagues, who may one day face an event similar to their event. They all demonstrated humility. They expressed admiration for the employees of their cities and appreciation for their peers who provided either direct aid or moral support.
They represent the values of city management and show how managers can rise to the occasion when extraordinary events occur. Their work is inspiring. Below are a few snapshots of lessons learned from city and county managers, whose interviews are featured in the Leading Edge Research report. Please share your general feedback. You can join in the discussion by joining the community or logging in here. You can also find out more about Emerald Engage. Visit emeraldpublishing. We will rather show that the Love Parade disaster resulted from the interaction of several contributing factors.
It is probably the first time that a detailed analysis can be performed with publicly available documents: not just investigation reports by public authorities 6 , 7 and the media, but also maps from Google Earth 8 and degree photographs, 9 videos accessible through YouTube, 10 documents released by Wikipedia 11 , 12 , 13 and Wikileaks, 14 and other sources. In some sense, this opens up a new age of public investigation.
However, to avoid misunderstandings, we would like to underline that our analysis focuses on the course of events and causal interdependencies among them, while they do not draw any conclusions regarding legal issues or personal or institutional responsibilities, which must be judged by other experts.
The remainder of this paper is structured as follows: Section 2 provides an overview of the situation before and during the Love Parade disaster. This includes a historical background, a description of the festival area including in- and outflows , and a timeline reconstructed from many video recordings.
Section 3 will analyze various factors contributing to the disaster, while Section 4 will focus on causal interdependencies and interaction effects. Section 5 discusses our findings and Section 6 concludes with lessons learned for the organization of future mass events. The novelty of this paper is four-fold: it concerns 1 the structured analysis of large amounts of publicly available video recordings of a disaster, 2 the interpretation of the disaster as a systemic failure where the interaction of various factors created a systemic instability, causing an overall loss of control , 3 a revision of common views about crowd disasters, and 4 the introduction of a scale reflecting the criticality of crowd conditions and proposed counter-measures.
The following section will try to give a short overview of the situation during the Love Parade in Duisburg and the planning beforehand. A large number of documents are now publicly available see collection of links.
It is certainly not possible but also not the purpose of this article to give a complete representation of materials. We will rather focus on the most relevant details in order to avoid a distraction of the reader from the main factors that have contributed to the disaster. The interested reader is invited to gain a more complete picture himself or herself, based on the media reports provided 21 , 22 , 23 and of several TV channels. In order to make an independent assessment possible, our own analysis will largely refer to authentic materials that are publicly accessible.
Videos of a subset of surveillance cameras are available until Timelines can be found in the endnotes. Many of these videos have been synchronized, 36 , 37 and some of them have been cut together in the form of multi-view videos documenting the course of events. The Love Parade is a popular electronic dance music festival in Germany that was first organized in Berlin in , and annually repeated in the same city until The events in and had to be cancelled because of funding problems and a coordinated opposition of political parties e.
The Love Parade in summer was again planned for Berlin, but the event was cancelled, since the Senate of Berlin did not issue the necessary permits on time. After negotiations with several German cities, it was then decided to move the Love Parade to the Ruhr Area, an agglomerate of major German cities, in the next years. The first of these events took place in Essen on August 25, , with 1.
In July , it was organized in Dortmund. The event, planned for Bochum, was cancelled due to security concerns, particularly as a critical situation had apparently occurred the year before. The chain of events underlying this disaster will be analyzed in the following sections.
The festival area of the Love Parade in was approximately , square meters large 21 and located in the area of a previous freight station of the city of Duisburg. For a degree view of the festival area and its surroundings see the endnotes.
In response to concerns from the regulatory authority that the area would be too small for the expected number of up to 1.
To overcome security issues seen by the regulatory authority there was some discussion to cancel the event overall , the organizer of the Love Parade decided to fence the whole festival area. However, there were still concerns that the standard safety requirements would not be met. The report argued that the festival area could be sufficiently well evacuated in an emergency situation.
Figure 1 gives an overview of the festival area. In the middle of that tunnel, there is the main ramp that leads to the festival area. The tunnel and the ramp together determine an inverse T-shaped geometry of in- and outflows.
The smallest overall diameter of the tunnels in the East and in the West was about 20 meters. However, the actual capacity was significantly lower than this due to the following factors see also Sec. The flow model of the organizer assumed numbers in Table 1. Illustration of the festival area and the ways to and from the area. Camera positions are shown as well as locations and events that are relevant for the analysis of this study. Note that the indicated timing of the police cordons as reconstructed from video recordings slightly differs from the police report, 6 but the differences are small.
According to Table 1 , between and the organizers expected an inflow of 90, and an outflow of 55, people, which could not have been handled by the wide ramp without the use of suitable crowd control. Problems had to be expected already for much smaller flow rates, as there were vehicles and a food stand as well as fences on the ramp, which must have reduced its capacity considerably.
This risk factor certainly had to be carefully considered by the crowd management concept. The festival area itself was apparently not overcrowded see caption of Table 1 and aerial photographs. To address this question, we will first present an expert opinion on the crowd disaster.
Then, we will summarize the course of events, and analyze the contributing factors in more detail. An expert report dated December 9, , which became public in February , 59 analyzes the implications of the flow model presented in Table 1. In the following, we summarize the essence of this report in our own words:. The organizational structure in particular, who takes what kinds of decisions must be fixed before the event. Particular attention must be paid to crowd management and communication loud speakers, signs, maps and other plans.
The division of responsibility should be regulated in the concept of the event. A mass event should not be approved, if it does not satisfy the applicable safety regulations.
Due to two triangular fence structures, 52, 62 which were apparently not shown in the maps, the effective width of the ramp was only According to the expert report, this implies a maximum safe flow of However, the maximum expected flow between and was , persons per hour, which would require a width of Therefore, at the Love Parade in Duisburg, problems with the in- and outflows and a critical accumulation of people had to be expected.
Therefore, injuries can easily happen. Due to a lack of suitable crowd control and guidance, visitors of the Love Parade in Duisburg could only see a narrow staircase as a possible emergency exit see Figure 1.
When trying to get there, the pressure towards the staircase increased and eventually triggered the crowd disaster. The analysis of the effective capacity of the main ramp suggests that problems on the ramp were foreseeable, and the question arises, why the obstacles were placed there.
However, a complete assessment should also consider the existence of the side ramp see Figure 1. Moreover, due to the applied access control, the flows on the main ramp did not reach the expected flows by far. This can be directly concluded from the fact that there was never any significant congestion between the two triangular obstacles defining the narrowest part of the ramp, before the flow was controlled in this area from on; this is clearly visible in the surveillance videos.
Queues of people did not form in the middle of the ramp, but rather at the upper end, where visitors were trying to enter the festival area. This, however, is not the location where the crowd disaster happened. Therefore, while one had to expect problems in the middle of the ramp where the triangular obstacles were located, the crowd disaster was actually not caused by those obstacles. The course of events that resulted in the crowd disaster involved many contributing factors, as we will show in the following.
The chronology presented in Table 2 is an abbreviated version of the timeline that was originally provided by the organizers of the Love Parade together with their documentary movie. Additional points will be discussed afterwards. Tables 3 and 4 present additional information that is relevant for a reconstruction of the causes of the crowd disaster. A time-ordered, geo-coded video link collection supplementing this paper allows the readers to gain an independent impression.
However, we would like to point out that the times provided on the videos or in the respective video portal may not always be exact. A synchronized video collection is now also available. After the occurrence of a disaster, it is natural to ask, who is responsible. In fact, after the Love Parade disaster, it seems that everybody was blaming everybody else: the visitors, the organizers, the police, the city of Duisburg.
What makes things difficult is that nobody is totally right and nobody is totally wrong: in the following, we will argue that it is the interaction of many contributing factors that caused the crowd disaster. Before we discuss the interaction of these factors, however, let us shed more light on some of them in separation.
While doing so, we will address a number of hypotheses regarding the cause of the crowd disaster, which have been formulated after the event. Given the many victims and pictures reminding of a war zone, some people first thought that a terrorist attack with explosives had happened. But what about the others? There certainly exist some instances of this kind such as the stampede in Baghdad on August 25, , due to spreading rumors of an imminent suicide bombing in the crowd, or the stampede in a Chicago night club triggered by rumors of a poisonous gas attack.
What evidence do we have for the Love Parade disaster in Duisburg? At first sight, one may think so, given that a number of visitors climbed over fences, up the pole, and on the container to reach the festival area.
However, as we will see, these activities started at a time when people on the ramp were already exposed to crowded conditions. Let us discuss this in more detail. The first problems with visitors overcoming fences were reported around Problems related to queues of people aggravate when queues are long and broad, so that little or no progress is visible.
In such situations, people will subconsciously reduce their distance eventually. However, it will also cause a compression of the crowd [ 47 ].
When the distance is small, there will be inadvertent body contacts, which can add up and cause unintentional pushing. Note that the transition from an acceptable situation with rare body contacts to a stressful situation with frequent contacts can happen quite abruptly, i.
People may interpret the situation as intentional pushing, which may trigger stress and aggression. At a certain density, it may also be required to push others away in order to be able to breathe [ 36 ]. If people have to wait long and are not informed about the reasons for this, they will become impatient and may eventually start to push intentionally because they assume that progress can be accelerated.
While most impatient pushing happens in the middle of the queue, the situation usually becomes most critical at the front of the queue but the people who push cannot see this, and they experience much less crowded conditions.
The situation is particularly bad behind bottlenecks. Such situations must generally be avoided. This also means that flow control is not a solution for every problem. It requires suitable designs and an adaptive operation. According to our assessment, it had to be expected that the access points would have to be opened and fences would eventually be overcome, given that the festival area and the inflow capacity were small in particular as the access was delayed by leveling works.
Waiting times often amounted to several hours, and access to entertainment, food, water, and toilets must have been quite limited outside the festival area. Nevertheless, the problems on the ramp were even more serious than at the access points.
They were related to the low inflow to the festival area see Table 1. An analysis of surveillance videos suggests that the floats i. After the crowd disaster, it was sometimes claimed that the floats even obstructed the inflow of arriving visitors. While the inflow never stopped completely before the cordons were established, the queue forming at the top of the ramp varied considerably over time.
It looks like the floats were slowed down by the dense crowd, which in turn obstructed the inflow of visitors, thereby creating an unfavorable feedback loop. The situation was particularly tense from to and from to ; as a consequence, the crowd manager asked for support by the police at or before.
However, already at i. In fact, it seems that the dangerous phenomenon of crowd turbulence see Sec. According to Table 3 , the first visitors used the narrow staircase at , and around the first people climbed the pole on the East side of the lower ramp area, to get up to the festival area. The main difference between these two modes of work is that outsourcing is much more like a typical occupation, with frequent communication and updates between the employer and worker. This back-and-fourth is a very different dynamic than the crowdsourcing dynamic, in which a task is given, then returned.
This outsourcing is possible in the Mechanical Turk marketplace, but the engine is not optimized for such activity and as a result most personal interactions take place via email. Suggestions and Lesson Learned We spent over 60 hours reading through Mechanical Turk Forum and Turkernation threads, reviews, and general discussions. Fig 9. Phase 2 submission quality vs. Phase 2 time logged vs. In Fig 9A we show the number of simulations which a participant ran before submitting one, compared with the quality of their submission.
A look at the graph shows that better results seem to weakly coincide with more simulation attempts, but not without significant outliers. In Fig 9B we show the amount of time for which a user was logged-in, compared with the quality of their submission. The time is measured in minutes. A look at the graph shows that better results seem to weakly coincide with more time spent, but not without significant outliers.
We gleaned a few notable ideas through this correspondence and from personal messages sent to us by crowd workers. More experienced crowd workers often become upset with newer crowd workers who will work for low paying requesters or those who are poor communicators.
We found that many crowd workers were suspicious of new requesters and high paying HITs. Turkopticon is a favored requester rating site for crowd workers. This site helps show what crowd workers believe is fair pay, how much communication is acceptable, the regular HIT acceptance speed, and the fairness of rejection rates. If a researcher is going to use Mechanical Turk, we suggest a few things: 1. Then just check the data you get back to make sure the code functions properly. Use Qualification HITs.
This will be the first HIT you release. All qualification tasks should be approved. That is, the crowdsourcer should mark the crowdworker's work as "approved". But, only crowd workers you want should be allowed to continue. This allows you to avoid rejecting the work of crowd workers, which creates ill will, while still not having to retain low quality crowd workers.
Crowd workers are very suspicious of potential rejection, as they don't want it to negatively impact their stats. However, if you have a massive amount of HITs to be completed quickly you may have to lower the qualifications. One option is to use the Masters qualification but only on the qualification HIT as in point 2. Then, after crowd workers are given your personalized qualification, you can take off the masters qualification knowing that all of your crowd workers are Masters and avoid paying the extra money that Mechanical Turk charges for using the Masters qualification.
But be cautious, we observed a strong negative feeling in the Mechanical Turk community with regards to this qualification. Use applicable social media, such as Turkernation to introduce yourself to workers and say that you will reward good workers and not reject sub-average workers unless they are obvi- ously not participating in good faith. Be watchful on such forums for questions from workers. Conclusions Given a sufficiently comprehensive tutorial and compensation, under certain circumstances, crowd workers can and will complete long and sometimes complex engineering tasks with comparable competence to that of formally trained persons.
This can be used to supply a signif- icant group of trainable assistants for experts to use to prescreen or check their own work. When giving such high difficulty tasks, one must be cognizant of the qualification limits placed on workers allowed to do these tasks.
Master crowd workers also have a higher return rate than non-Masters. The data in this experiment encourages the use of crowd workers as trainable assistants for complex engineering tasks.
Tasks, such as surveys, and data analysis can be done quite easily on the Mechanical Turk with fast and accurate results if one is using the correct qualifications and has a sufficiently well-designed tutorial.
If one asks crowd workers to use complicated, unfamiliar tools, this may result in significant increase in the need for communication between the crowdsourcer and the crowdworkers, but such tasks can be done. Future Work Given the outcome of this experiment, we are interested in studying the effectiveness of crowd workers on other complex engineering tasks.
One option is to consider pushing the complexity of the tasks even more, such as by asking crowd workers to design their own buildings using Google Sketchup or a similar program and then analyzing them.
However, this runs the risk of further pushing the experiment into outsourcing, rather than crowdsourcing, and only gaining a limited pool of crowd workers, which would have to be trained even more and provides less of a basis for consensus. If, instead, we wish to remain in crowdsourcing, we see three main avenues of exploration. We can explore other high-complexity tasks.
For example, we can ask crowd workers to help in the processing and analysis of post-disaster reconnaissance images gathered by apps such as CyberEye [21].
We could also investigate the internal dynamics of crowdsourcing itself, by studying the effectiveness of introducing requesters on networks such as Turkernation, Mechanical Turk Forum, and CloudMeBaby. One potentially interesting avenue of investigation is a hybrid model which includes auto- mated computers, crowd workers and experts.
In one particular design, we can have computers automatically perform many simulations, of fairly coarse quality over a range of parameters, on designs proposed by experts.
Then, we can ask crowd workers to evaluate the quality of those simulations. Using the crowd results, we can further refine the simulations, producing another set of simulations of finer detail over the parameter ranges garnering crowd consensus. We can then repeat this until we have a simulation of the precision desired, which could be fur- ther refined by an expert. We would argue that one of the most interesting aspects of this entire study is that we have begun to more greatly appreciate the considerable potential of crowdsourcing, both as a subject of study and as a means of accomplishing more and more complex tasks.
It is our hope that, as we begin to further understand the potential of this tool, we can more effectively deploy it for some complex, real world, engineering problems. Supporting Information S1 Dataset. Response data for grad student group 1 to Phase 1 questions.
Some questions are removed due to personally identifying data. CSV S2 Dataset. Response data for grad student group 2 to Phase 1 questions. As explained in the paper, this data is in a slightly different format than the previous dataset.
But the image names can be used to correspond to the other data sets. Aggregate scores of all participants. This file contains the percentage "scores" relative to ground truth of all participants who completed all tasks in phase 1 up to 12 "Unable to determine" or blanks allowed. NT—Non- master Turker.
U1—Grad student group 1. U2—Grad student group 2. CSV S4 Dataset. Response data for master turkers to Phase 1 questions. There were two groups of master turkers surveyed, one in , and one in The beginning of each group is marked.
CSV S5 Dataset. Response data for non-master turkers to Phase 1 questions. There were two groups of non-master turkers surveyed, one in , and one in CSV S6 Dataset. Demographic and motivational data for crowd workers who completed an extended demographic survey. For survey questions, all column headings explain the questions asked. For survey questions, options to not answer or answer "other" were given. Also contained in this file are aggregate statistics for crowd workers from the 5 survey questions which were asked of all participants.
CSV S7 Dataset. Summary data for the submission quality for Phase 2. CSV S8 Dataset. Aggregate summary of the interactions of users with the Virtual Wind Tunnel during phase 2, for users who made a submission. Images for the users to evaluate in phase 1. Set of folders containing the key images, extracted from the simulations submitted by users for phase 2. The numbers for each folder corresponds to the user number in phase2reports. The Virtual Wind Tunnel does not generate images unless a user asks for a particular image.
On a few occasions, users did not even look at certain graphs, so those graphs were not generated. Such non-inspected graphs are not present here. Note: the x-axis label in the wake stream velocity graph, in Phase 2, due to a typo, indicated a scaling factor that was not actually applied.
Charles Jhin assisted with the statistical analysis. Performed the experiments: MS PS. Provided expert support in fluid mechan- ics: DW AK.
0コメント