Se ha denunciado esta presentación.
Utilizamos tu perfil de LinkedIn y tus datos de actividad para personalizar los anuncios y mostrarte publicidad más relevante. Puedes cambiar tus preferencias de publicidad en cualquier momento.
Practical Pilot Amsterdam
Putting 20 years of research to practice
Prof. dr. Serge Hoogendoorn, Delft University of Techno...
Efficient self-
organisation
Capacity-drop
and start-stop
waves
Blockages, gridlock and
inefficient choice
behaviour
“Ther...
Why do we need to control traffic at all?
• When traffic is dilute, there is little need to intervene to improve throughpu...
Example phenomena: capacity drop
• Road capacity changes after the on-set of congestion
• Capacity drop magnitude depends ...
Example phenomena: capacity drop
• Road capacity changes after the on-set of congestion
• Capacity drop magnitude depends ...
Example: isolated ramp metering
•With ramp-metering, we can delay capacity drop
•Simple feedback scheme may do the trick!
...
No control Isolated control
Example: isolated ramp metering
•With ramp-metering, we can delay capacity drop
•Simple feedba...
Example phenomena: 

Start-Stop Waves (or Wide
Moving Jams)
• Busy traffic is inherently unstable
• Small disturbances inc...
Example phenomena: 

Start-Stop Waves (or Wide
Moving Jams)
• Busy traffic is inherently unstable
• Small disturbances inc...
Example: using Specialist to remove wide-moving jams
•Specialist removes these jams by means of speed-limit control
•By re...
•How to increase the effectiveness of the local controllers considered?
• In examples shown effectiveness of local control...
• Recent MPC-based network-wide control approaches with colleagues at Swinburne
(with Tung Le, Hai le Vu, and Han Yu) allo...
In the Field Operational Test “Praktijkproef Amsterdam” 

TU Delft is developing operational control methods 

for coordin...
• Bottleneck is detected (or predicted)
• Nearby local measure starts resolving problem
by deploying local control approac...
• Master ramp controller:
• Slave ramp controller:
with:
and:
• Intersection controller (adaptation of vehicle-
response c...
Before field deployment, system was analysed
•Tuning of control parameters
•Test in simulation environment (macroscopic mo...
Which monitoring and control functions are needed?
Towards a functional architecture and design
LCE (RM algorithm)
Supervi...
Distinguished monitoring and control functions
The Functional Architecture of the PPA system
Parameter
Estimator
Storage S...
Data collection for Behavioral Modeling - ICEM 2012
• PPA substantially improves freeway throughput (daily reduction delay...
Data collection for Behavioral Modeling - ICEM 2012
• PPA substantially improves freeway throughput (daily reduction delay...
Data collection for Behavioral Modeling - ICEM 2012
• PPA substantially improves freeway throughput (daily reduction delay...
• Overall conclusion: advanced INM is feasible
• Careful and thorough analysis of the outcomes
has led to a number of impo...
• Bottleneck inspector used advanced data mining techniques and traffic flow
modelling to predict if in the next 3-5 minut...
• Approach cannot distinguish between on-ramp bottleneck, spill back from
downstream bottleneck, or congestion due to inci...
• Approach cannot distinguish between on-ramp bottleneck or congestion due to
incidents downstream
• Control strategy is n...
• We determined that for each vehicle we held
back for 1 min, we save 2 veh-min delay on
the freeway
• This means that for...
Transitions in Traffic Management
Combining road-side and in-car approaches
Steps towards phase 2 of the pilot
2012 2013 2...
• Case with 2 OD pairs and perfectly informed traveler from A to B having 2 options
• Intelligent intersection controller ...
• Case with 2 OD pairs and perfectly informed traveler from A to B having 2 options
• Intelligent intersection controller ...
• Data fusion is in-car and road-side data is not the only opportunity!
• Anticipatory network management: optimise traffi...
• Important note: modification of Specialist control algorithm required to
incorporate changes in flow operations for diff...
• Presented approach has been successfully generalised to other bottlenecks
(including spill-back), including those on urb...
Efficient self-
organisation
Faster =
slower effect
Blockades and
turbulence
“There are serious limitations to the self-or...
Engineering the future city.
Crowd Management Dashboard
- Generate and visualise information to support
operations and ana...
REAL-TIME DATA COLLECTION USING MULTIPLE SOURCES
Counting cameras
collect flow
information
Wifi sensors track
smartphones
...
EXAMPLE PILOT STUDY RESULTS
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
11 12 13 14 15 16 17 18 19
Dichtheden)Sumatrakade
dichtheid2(ped...
RESULTS OF PILOT
Future research steps…
- Validated and generic system that will work for
‘any’ outdoor (& indoor) event /...
Practical Pilot Amsterdam
Putting 20 years of research to practice
Prof. dr. Serge Hoogendoorn, Delft University of Techno...
Próxima SlideShare
Cargando en…5
×

IPAM Hoogendoorn 2015 - workshop on Decision Support Systems

306 visualizaciones

Publicado el

Presentation during IPAM workshop in Los Angeles where I shared the results of the Practical Pilot Amsterdam (a pilot of Integrated Network Management in Amsterdam), the lessons learnt and the plans for the next phase.

Publicado en: Ciencias
  • Sé el primero en comentar

IPAM Hoogendoorn 2015 - workshop on Decision Support Systems

  1. 1. Practical Pilot Amsterdam Putting 20 years of research to practice Prof. dr. Serge Hoogendoorn, Delft University of Technology
  2. 2. Efficient self- organisation Capacity-drop and start-stop waves Blockages, gridlock and inefficient choice behaviour “There are serious limitations to the 
 self-organising capacity of traffic systems” Increasing network load Reduced production of network Why do we need to control traffic at all? • When traffic is dilute, there is little need to intervene to improve throughput • However, when traffic loads become higher, phenomena occur that reduce there efficiency of network operations dramatically (Network Fundamental Diagram)
  3. 3. Why do we need to control traffic at all? • When traffic is dilute, there is little need to intervene to improve throughput • However, when traffic loads become higher, phenomena occur that reduce there efficiency of network operations dramatically (Network Fundamental Diagram) BlockadesCapacity drop Uneven distribution of traffic Inefficient choice behavior Improve traffic distribution Prevent blockades Improving throughput Reducing inflow Reduced performance of network in case of overloading (Network Fundamental Diagram) Solution directions or golden rules 
 of traffic control and management (how to deploy measures)
  4. 4. Example phenomena: capacity drop • Road capacity changes after the on-set of congestion • Capacity drop magnitude depends on severity of congestion caused by the bottleneck (i.e. the speed in the queue) locatie(km) tijd (u) 7 7.5 8 8.5 9 9.5 36 38 40 42 0 50 100 7 7.5 8 8.5 9 9.5 0 20 40 60 80 100 120 tijd (u) snelheid(km/u) 5 7 8 9 5 7 8 9 7.5 8 8.5 9 9.5 −4400 −4300 −4200 −4100 −4000 −3900 −3800 tijd (u) N(t,x) q0 = 3700 7 8 9 7 7.2 7.4 7.6 7.8 8 −5050 −5000 −4950 −4900 −4850 −4800 −4750 tijd (u) N(t,x) q0 = 4000 7 8 9 
 • Figure shows on-ramp bottleneck 
 on Dutch A9 motorway • Congestion sets in at around 7:30 
 and lasts until 9:30 • Middle line in bottom figure shows slanted cumulative curve, slope of which is the flow minus a reference flow (3700 veh/h) • Pre-queue flow = 4200 veh/h, post-queue flow = 3750 veh/h • Capacity drop is (at least) 11%, but can be higher depending on the congestion severity
  5. 5. Example phenomena: capacity drop • Road capacity changes after the on-set of congestion • Capacity drop magnitude depends on severity of congestion caused by the bottleneck (i.e. the speed in the queue) • Picture shows relation between speed in the queue upstream and the size of the capacity drop • Capacity drop goes up to 30% is vehicles in queue as standing still (what type of queue is this?) • Preventing or removing congestion will enable flow to operate in a more efficient pre-queue state
  6. 6. Example: isolated ramp metering •With ramp-metering, we can delay capacity drop •Simple feedback scheme may do the trick! •Maintaining high-capacity regime causes substantial improvements in total network delay • Example: dynamic ALINEA-type algorithms steers downstream density to optimal value (generally the critical density) • Test in Amsterdam shows 8% increase in outflow compared to no control case • Data analysis shows that this leads to an 1:2 gain (freeway):loss (ramp) ratio, which we can further improve by better choosing the target value qramp(t + 1) = qramp(t) + K · (⇢⇤ ⇢(t)) qramp(t) ⇢(t)
  7. 7. No control Isolated control Example: isolated ramp metering •With ramp-metering, we can delay capacity drop •Simple feedback scheme may do the trick! •Maintaining high-capacity regime causes substantial improvements in total network delay • However, metering is terminated if on-ramp queue spills over to urban network • Limited space to buffer traffic yields large impact on effectiveness of ramp-meters (average up time in The Netherlands = 8 min) bufferspace depleted
  8. 8. Example phenomena: 
 Start-Stop Waves (or Wide Moving Jams) • Busy traffic is inherently unstable • Small disturbances increase in amplitude as they move from vehicle to vehicle and may eventually cause vehicles to stop • Start-stop wave is born and will remain until inflow is lower than outflow for sufficiently long period • Note that outflow of wave is around 30% less than the free road capacity • Waves can persist for a very long period • From the road-user they are unexpected and hence unsafe
  9. 9. Example phenomena: 
 Start-Stop Waves (or Wide Moving Jams) • Busy traffic is inherently unstable • Small disturbances increase in amplitude as they move from vehicle to vehicle and may eventually cause vehicles to stop • Start-stop wave is born and will remain until inflow is lower than outflow for sufficiently long period • Note that outflow of wave is around 30% less than the free road capacity • Waves can persist for a very long period • From the road-user they are unexpected and hence unsafe
  10. 10. Example: using Specialist to remove wide-moving jams •Specialist removes these jams by means of speed-limit control •By reducing inflow into wave sufficiently and sufficiently long • After tuning, we had 2.8 activations per day resolving the jam in 72% of the cases • What limits number of activations?
  11. 11. •How to increase the effectiveness of the local controllers considered? • In examples shown effectiveness of local controllers is limited due to limited buffer space (e.g. on-ramp, roadway stretch along which speed limit is applied) • Effectiveness can be increased by using space in network elsewhere (buffering) • This calls for coordination of available traffic management measures (freeway and urban), which has been a research subject for a long time (e.g. the EU DACCORD project starting in 1995, following the DRIVE / DRIVE II project) • Only few successful practical pilots have however been conducted, due to 
 1) complexity of many of the (theoretical) approaches proposed or 
 2) ineffectiveness simple (scenario-based) approaches applied in Dutch practise • Objective of Practical Pilot Amsterdam (PPA) is to get practical experience with effective Integrated network management… Increasing effectiveness by coordination Use ‘control space’ elsewhere in the network But do it wisely!
  12. 12. • Recent MPC-based network-wide control approaches with colleagues at Swinburne (with Tung Le, Hai le Vu, and Han Yu) allows including main phenomena observed in traffic flow (capacity drop, spillback, congestion dynamics) • Formulation of the MPC problem as a LQ controller with inequality constrains can be solved efficiently using dedicated solvers (CPLEX) • Contributions of Le et al (2013), Yu et al (2015) on this subject available (send me an email) Intermezzo: what we could not sell… Model Based Predictive Control approaches Computation time is not necessarily the problem…
  13. 13. In the Field Operational Test “Praktijkproef Amsterdam” 
 TU Delft is developing operational control methods 
 for coordinated control (planned in 2013) Expecting a reduction in Vehicle Loss Hours of over 30%
 due to using recent insights in dynamics and control N200 N220 RAI S112 S100 S116 S114 A9 A10 A10 A8 S102 ArenA A5 Introducing the Praktijkproef Amsterdam •Phase 1 focusses on A10W and s102 (this presentation) •Phase 2 considers other locations + integration road-side / in-car (monitoring) •Phase 3 considers integration road-side / in-car (control / actuation) A10W s102 Coentunnel Phase 1 (July 2013): road-side • Coordinated ramp meters A10W • Coordinated intersection control s102 • Coordination A10W and s102 Phase 1: in-car (separate!) • Different trials with (mostly) in FCD data collection, in-car information provision and guidance • Regular and event conditions Coentunnel had capacity drop op 13% Approach is based on preventing drop by limiting flow towards bottleneck Spill-back and grid-lock isprevented as much as possibleon intersections and connections
  14. 14. • Bottleneck is detected (or predicted) • Nearby local measure starts resolving problem by deploying local control approach • The local measure (Master) is supported by measures elsewhere (Slaves): - Slave ramp-meters limit flow towards bottleneck causing reduced flow on freeway - Slave traffic controllers limit flow to ramp - Slave VSL control reduces flow towards the bottleneck (using VMS or in-car devices) • Ensure efficient distribution of queues over available buffers using feedback mechanisms (generalisation of HERO principe) Increasing effectiveness by coordination Use ‘control space’ elsewhere in the network But do it wisely! Bottleneck Relative space of Slave buffer = relative space of Master buffer Note that similar approaches work on different types of bottlenecks
  15. 15. • Master ramp controller: • Slave ramp controller: with: and: • Intersection controller (adaptation of vehicle- response controller): with: • VSL control (not implemented) based on changing region location Increasing effectiveness by coordination Use ‘control space’ elsewhere in the network But do it wisely! Bottleneck Relative space of Slave buffer = relative space of Master bufferqramp(t + 1) = qramp(t) + K · (⇢⇤ ⇢(t)) qramp(t + 1) = qramp(t) + K1 · e(t) + K2 · e(t) e(t) = e(t) e(t 1) Gext j (t + 1) = Gext j (t) + k1 · e(t) + k2 · e(t) e(t) = srel Master(t) srel Slave(t) e(t) = srel Master(t) srel j (t)
  16. 16. Before field deployment, system was analysed •Tuning of control parameters •Test in simulation environment (macroscopic models, microscopic simulation with Vissim) Eigenwaarde shows speed of convergence (blue is better) Example of controller behaviour for three choices of (K1,K2) • Tuning by naive trial and terror? We can do better! • Using mathematical systems’ theory to study convergence behaviour via max eigenvalue for gains K1 and K2 (TRB paper 2015): |e(t + 1)| ⇡ max · |e(t)| max
  17. 17. Which monitoring and control functions are needed? Towards a functional architecture and design LCE (RM algorithm) Supervisor A10W Bottleneck Identifier Queue Estimator Parameter Estimator Subnetwork Supervisor Supervisor T1 Light (ST1L)Supervisor T1 (ST1) (RTNR controller) • Detect or predict occurrence of a bottleneck (3-5 minutes ahead) • Determine ramp-metering control target (e.g. critical density) • Estimate queue lengths on-ramps and urban arterials • Determine metering rate at the Master ramp • Determine which buffers to use for coordination support • Determine metering rate of the Slave ramp meters • Determine extension green Traffic Controllers for rel. buffers
  18. 18. Distinguished monitoring and control functions The Functional Architecture of the PPA system Parameter Estimator Storage Space Identifier LCE (RM algorithm) LCE (Traffic Control) LCE (Traffic Control) Supervisor A10W Supervisor T1 Light (ST1L) Supervisor T1 (ST1) (RTNR controller) Subnetwork Supervisor Ramp-meter Congestion Estimator Bottleneck Identifier Bottleneck Identifier Level of Service Indicator Queue Estimator FractionEstimator Logical Monitoring Units Logical Control and 
 Supervisor Units
  19. 19. Data collection for Behavioral Modeling - ICEM 2012 • PPA substantially improves freeway throughput (daily reduction delays of 300 veh-h) • Increase delays on urban arterials, can be reduced by about 80% by better tuning and configuring system and by use of in-car technology / data fusion • Expected gross impact of 220 veh-h on daily basis (around 1 million Euro/year)
  20. 20. Data collection for Behavioral Modeling - ICEM 2012 • PPA substantially improves freeway throughput (daily reduction delays of 300 veh-h) • Increase delays on urban arterials, can be reduced by about 80% by better tuning and configuring system and by use of in-car technology / data fusion • Expected gross impact of 220 veh-h on daily basis (around 1 million Euro/year) • Average total delay in evening peak with and without PPA Total delay (in hours) per minute No PPA With PPA Tijd (u) Ruimte(hm) Snelheidscontour (07−Nov−2013) 15:30 16:00 16:30 17:00 17:30 18:00 2.6 2.7 2.8 2.9 3 3.1 3.2 x 10 4 0 20 40 60 80 100 120 Coordination allowed us to meter 3-5 times longer than with local metering
  21. 21. Data collection for Behavioral Modeling - ICEM 2012 • PPA substantially improves freeway throughput (daily reduction delays of 300 veh-h) • Increase delays on urban arterials, can be reduced by about 80% by better tuning and configuring system and by use of in-car technology / data fusion • Expected gross impact of 220 veh-h on daily basis (around 1 million Euro/year) Picture shows locations where delays improved or got worse Correction for demand levels for considered days
  22. 22. • Overall conclusion: advanced INM is feasible • Careful and thorough analysis of the outcomes has led to a number of important lessons, amongst which are the following: 1. The value of prediction and good monitoring approaches 2. Bottleneck diversity 3. Buffer effectiveness • These lessons are effectuated in phase 2 of PPA, where also integration with in-car data collection will be considered… • Let’s take a closer look at these improvements… Lessons learnt from phase 1 of the pilot Value of carefully assessing impacts Steps towards phase 2 of the pilot
  23. 23. • Bottleneck inspector used advanced data mining techniques and traffic flow modelling to predict if in the next 3-5 minutes congestion would occur, after which the system would start intervening • Due to high change of false positives, the system generally started too soon (our data analysis shows 34 minutes too soon) with reducing inflow / filling up buffers • Overall result: expected reduction delays on urban network of around 40% due to - removing unnecessary delays on on-ramp and urban arterials - increasing effectiveness ramp metering (since one buffer space was needed, part was already used) •In phase 2 we therefore refrain from use of prediction (for this type of bottleneck), instead opting for timely detection of bottleneck instead Lessons learnt from phase 1 of the pilot The value of prediction and good monitoring Steps towards phase 2 of the pilot Can models predict on-set ofcongestion with a sufficientlyhigh probability given highvariation is demand and supply?How to effectively use models?
  24. 24. • Approach cannot distinguish between on-ramp bottleneck, spill back from downstream bottleneck, or congestion due to incidents • Control strategy is not dependent on bottleneck type (also holds for local meas.) • In case of incident, normal strategy will not be effective in removing problem •In phase 2, bottleneck inspector
 will distinguish between 
 different types of bottlenecks 
 and control strategy will be 
 dependent on it • Example shows different bottlenecks
 that occur (on-ramp queue, merge
 queue, wide-moving jam) Lessons learnt from phase 1 of the pilot Bottleneck diversity Steps towards phase 2 of the pilot
  25. 25. • Approach cannot distinguish between on-ramp bottleneck or congestion due to incidents downstream • Control strategy is not dependent on bottleneck type • In case of incident, normal strategy will not be effective in removing problem •In phase 2, bottleneck inspector
 will distinguish between 
 different types of bottlenecks 
 and control strategy will be 
 dependent on it • Example shows different bottlenecks
 that occur (on-ramp queue, merge
 queue, wide-moving jam) PPA phase 2 Bottleneck diversity • Animation shows the GUI developed for phase 2 • Approach can deal with four types of congestion using separate control strategies (weaving areas, infrastructure bottleneck, on- ramp bottleneck and wide-moving jam) • Animation shows different types occurring during the morning peak
  26. 26. • We determined that for each vehicle we held back for 1 min, we save 2 veh-min delay on the freeway • This means that for buffers with less than 50% of the vehicles actually traveling to the bottleneck the delay we incur is actually larger than the improvement we make • Large share of the buffers were not effective at all! Additional reduction of urban delays with about 42% •Phase 2 will involve the use of FCD data to determine the actual shares and include / exclude buffers dynamically Lessons learnt from phase 1 of the pilot Buffer effectiveness Steps towards phase 2 of the pilot In general, effective buffers are only found inproximity of bottleneck
  27. 27. Transitions in Traffic Management Combining road-side and in-car approaches Steps towards phase 2 of the pilot 2012 2013 2014 2015 2016 IN-CAR TRACK ROADSIDE TRACK INTEGRATION TRACK PHASE 2PHASE 1 PHASE 3 PRACTICAL TESTS FULL SCALE APPLICATION GO/ NO GO ROADSIDE GO/ NO GO INTEGRATION GO/ NO GO INTEGRATION First steps in integrating road- side and in-car approaches are focused on monitoring: • Determining buffer fractions • Data fusion to improve queue length estimations • Testing if we can do with less loops on the freeway What about phase 3?
  28. 28. • Case with 2 OD pairs and perfectly informed traveler from A to B having 2 options • Intelligent intersection controller optimises locally capacity use at intersection evenly (equal delays for both directions) • Most travellers choose route 2 under normal circumstances • Event (incident) occurs; flow conditions route 2 worsen (from 120 to 80 km/h) A Simple Quiz Anticipating Intelligent Intersections Need of including traveler response into the design A C D B Route 1 Route 2 • Route 1 becomes more attractive for A-B travellers (more people choose route 1) • Intersection controller assigns more capacity to A-B travellers due to increased demand • Route 1 becomes more attractive for A-B travellers; situation worsens form C-D travellers Quiz: what happens assuming 
 fully informed road users?
  29. 29. • Case with 2 OD pairs and perfectly informed traveler from A to B having 2 options • Intelligent intersection controller optimises locally capacity use at intersection evenly (equal delays for both directions) • Most travellers choose route 2 under normal circumstances • Event (incident) occurs; flow conditions route 2 worsen (from 120 to 80 km/h) A C D B Route 1 Route 2 A Simple Quiz Anticipating Intelligent Intersections Need of including traveler response into the design perc. choice route 1 trafficcontroller total delays 0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4 0.5 1 2 3 4 5 6 x 10 5 Total delays in system increase with about 30% Start 1 2 3
  30. 30. • Data fusion is in-car and road-side data is not the only opportunity! • Anticipatory network management: optimise traffic management and control measures anticipating on the impacts it has on travel decision • Theoretical studies show that we can get very close to system optimum • Future use of the ‘car’ as a traffic management measure Start 1 3 A Simple Quiz Anticipating Intelligent Intersections Need of including traveler response into the design Conclusion: intelligentintersection control can bedesigned to achieve goodnetwork performance if itanticipates correctly ontravel responses due toadvanced information services No anticipation on traveler response
  31. 31. • Important note: modification of Specialist control algorithm required to incorporate changes in flow operations for different penetration levels Cooperative version of Specialist Application of shockwave suppression algorithm V2I approach allows application at different penetration rates
  32. 32. • Presented approach has been successfully generalised to other bottlenecks (including spill-back), including those on urban roads • Work presented is aimed at developing methods that can be applied in practise, which has implications for transparency and allotted complexity • Questionable if “adding complexity” will yield major improvements: - In practise, relation between buffers and bottleneck becomes weak quickly (in a spatial sense) so controlled subnetworks are relatively small - Stochasticity in demand and supply makes it very hard to predict moment of congestion on-set and thus limits applicability of prediction models (not saying that they are of no use!) • Queue estimation (with loops) turned out to be a major challenge, consider use of alternative monitoring technology or other control variables (which?) • Concepts are not only applicable to vehicular traffic… Closing remarks Lessons learnt and future steps And some other considerations
  33. 33. Efficient self- organisation Faster = slower effect Blockades and turbulence “There are serious limitations to the self-organising abilities
 of pedestrian flow operations” Toenemende belasting verkeersnetwerk Reduced production of pedestrian network Flow operation ‘self-degradation’ also occurs in pedestrian flow • Dilute pedestrian traffic organises itself very efficiently (bi-directional / crossing flows) • When network loads increase, phenomena occur that severely reduce production • Need for crowd management!
  34. 34. Engineering the future city. Crowd Management Dashboard - Generate and visualise information to support operations and analysis (during the day, after each day), by… - Collecting real-time data about pedestrian flows (travel times, crowdedness, route choice, background visitors) - Combining these multiple sources to get best picture of current situation, overcoming limitations of individual data sources
  35. 35. REAL-TIME DATA COLLECTION USING MULTIPLE SOURCES Counting cameras collect flow information Wifi sensors track smartphones Some visitors equipped with GPS devices Social Media data (social-media activity per area, visitor characteristics) Furthermore… - Cameras to verify if dashboard information is indeed correct - Areas photographs from balloon
  36. 36. EXAMPLE PILOT STUDY RESULTS 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 11 12 13 14 15 16 17 18 19 Dichtheden)Sumatrakade dichtheid2(ped/m2) 0 5 10 15 20 25 30 11 12 13 14 15 16 17 18 19 Verblijf) en+looptijden+Veemkade verblijftijd looptijd
  37. 37. RESULTS OF PILOT Future research steps… - Validated and generic system that will work for ‘any’ outdoor (& indoor) event / regular situation - Trial prediction capabilities to warn crowd managers in advance for risky situation - Visitor information provision (signs, Smartphone) - Provide advice to crowd managers on which measures to deploy - Support optimal use of available infrastructure for safe and comfortable stay in city
  38. 38. Practical Pilot Amsterdam Putting 20 years of research to practice Prof. dr. Serge Hoogendoorn, Delft University of Technology Questions

×