Modelling traffic disruption impact: Deep Learning & simulation
This PhD project will focus on the complex problem of predicting the impact of traffic disruptions in large cities using advanced artificial intelligence algorithms and evaluating the best response plan that traffic authorities can make by synergising traffic simulation modelling of various response scenarios.
Today, Intelligent Transportation Systems (ITS) are an integral element of any transport network as they provide monitoring and control, ensuring travel safety, increasing road efficiency, reducing travel time, etc.
However, despite all efforts of ITS systems to optimise congestion and ease traffic, accidents can appear anywhere at any point in time, affecting traffic congestion and planned clearance times. Traffic disruptions can be recurrent (daily congestion patterns) or non-recurrent (instantaneous traffic incidents caused by human errors/bad transport design planning, etc.).
Important tasks to bring this project to fruition include:
- Detection of traffic incidents, traffic incident severity and incident duration prediction (based on traffic flow parameters and incident log datasets)
- Spatial-temporal incident impact estimation (including impact map of adjacent transport elements and estimation of spatial-temporal lifecycle of disruption)
- Simulation-based modelling for response plan and incident clearance strategy.
The solution would improve traffic centres decision making by automatic response plan recommendations.
Today, Intelligent Transportation Systems (ITS) are an integral element of transport networks in modern cities. These systems provide monitoring and control of the transport system, ensuring safety, increasing efficiency, reducing travel time, reducing air emissions and thus having a significant impact on the economy and health of the city population.
However, despite all effort for ITS systems to optimise congestion and ease traffic, accidents can appear anywhere at any point in time. Traffic disruption is an unwanted specific effect of severe congestion, which can be observed in any transport systems. Disruption can be recurrent (characteristic to highly congested transport networks) or nonrecurrent (rarely observed, traffic incident). With recent advances in information technologies, new possibilities emerge for incident classification, spatial-temporal analysis and traffic simulation.
In Australia, there is currently a lack of incident management and response plans solutions and the majority of transport management centres make decisions based on operational experience of staff members rather than data-driven lessons learned. There is a true lack in adopting more advanced solutions that can make use of any existing sources of information and modern data-driven techniques such as machine learning /deep learning.
Also in the literature, there is a lack of exploring various all-modelling capabilities combining transport modelling and data-driven solutions. In the paragraph below, we provide the context and the literature review for each of the major steps of this project.
Stage 1. Incident Detection, Incident Duration Modelling, Incident Severity Prediction
Context: Most current studies rely on methods for classification and clustering of traffic conditions for doing the incident detection. However, there are very few studies on traffic incidents involving methods for detecting anomalies (such as one-class SVM, isolation forests, etc).
Non-recurring traffic incidents are rare and unusual in nature and therefore the detection of a traffic incident can be assessed as the task of detecting anomalies in traffic. By relying on anomaly detection methods, the incident detection system can be adapted to previously unseen situations. Thus, classification and evaluation of the applicable anomaly detection methods in comparison to well-established classification and regression methods will be carried out.
Also, road situations detected as anomalous can be extremely valuable for further investigations in the duration of a freshly reported accident.
Incident duration distribution has been modelled as log-normal1 and more recently as log-logistics distribution2,3. Log-logistic model has been used more extensively and found to have better goodness-of-fit than lognormal distribution.
Also, there are various hazard-based models of traffic incident duration4, 5 which employ a hazard function to describe the conditional probability that an incident will end during any particular time interval given that it already lasted until the beginning of the interval. Recent studies also involve multicomponent log-logistic models. The authors in6 describe a g-component log-logistic model and7 competing risk mixture model which incorporates a multinomial log-logistic model.
An actual distribution estimation can give only approximate information on traffic incident duration. More than that, incident duration distribution is found significantly dependent on incident case parameters (e.g. day/night)8. Also, the duration of the incident may be affected by the chosen method of incident clearance7. Other authors9 found that some incident parameters found to be important factors with different contribution to different types of accidents, including weather condition, traffic density, time period, and incident location.
Also, road factors found to be affecting each one of 4 incident types (rear-end, side wipe, collision with fixtures and rollover) in a different way. These findings draw incident duration distribution estimation as a complex problem dependent on several traffic flow and incident parameters.
Incident duration can be modelled in terms of spatial relations (geometric placement of adjacent lanes, angle of adjacency, different parameters of street lanes, including speed limits).
Recent studies rely on reported incident parameters10, 11, but road topology can also play significant role in estimation of the incident probability (e.g. poorly designed junction, wrongly imposed speed limits). Approximately 5% of the road junctions are the site of 50% of the accidents in the city of London.12
Thus, it seems reasonable to analyse incident duration and probability with consideration of the road topology. The task of predicting the duration of an incident is usually solved by using Machine Learning methods. Among these methods tree-based methods13, fuzzy logic14, Bayesian networks15, artificial neural networks16, 17, and recently GBDT have been revealed to be a better performing method for incident duration prediction18. Gaussian process regression and artificial neural networks were found to outperform tree methods and SVM in incident duration prediction11.
Also, estimation of incident duration can be reduced to the classification method10. To do this, a specific threshold for the duration is set and a prediction is made whether the incident will last longer than a specified time. Artificial neural networks show high average accuracy for prediction of 4 types of incident severity relying on data on the state of the road (lane, condition of the roadway, weather, light, etc.), time and date.
Overall accuracy between death, severe, moderate and minor severity accidents was found to be 69-72%17. Recent studies in machine learning involve interpretable models. Bayesian networks can produce interpretable models for incident injury severity prediction19.
Bayesian networks also outperform regression models in incident severity prediction (involving three severity indicators: number of fatalities, number of injuries and property damage)20. Interpretability is not strictly limited to tree models and by using knowledge distillation one can extract tree rules from different prediction models (e.g. Bayesian network21). It allows to represent the model as an interpretable decision tree and estimate feature importance.
Stage 2. Spatio-temporal incident modelling
Context: By using the spatial-temporal forecast, it is possible to produce an estimation of the road situation development, which can be used in planning of a strategy to eliminate the incident. Also, this kind of forecasting allows to produce incident affectability map for every component of the city road network.
This kind of data can be used in road planning decisions to reduce incident impacts on road network in long-term and to reveal road elements which are the most sensitive to traffic incidents (e.g. produce wide-spread or long-term congestion). Spatial-temporal incident modelling (e.g. impact forecasting) is prior for the next stage – incident response modelling.
Figure 1 – Visualisation of spatial-temporal incident impact22
Spatial-temporal impact can be estimated using different approaches of traffic simulation. There are 3 groups of traffic simulation models:
- Microscopic – traffic network is simulated on the level of individual agents (car, pedestrian), relying on rules of movement and interaction (including lane change, acceleration). These models include Car-following models (which relies on real driving behaviour such as keeping a “safe distance” from the leading vehicle23), such as Gipps model, Intelligent driver model, etc. All of these methods require a lot of computational resources and they follow a dynamic traffic assignment modelling.
- Mesoscopic – traffic is represented by interactive groups of traffic entities moving together in the network and taking turns by following proportional turning percentages. They incorporate the notion of time and can be used in a hybrid approach for specific parts of the network, together with microscopic simulation modelling.
- Macroscopic – traffic is simulated on the level of traffic flow on road segments and follows a static traffic assignment approach (where the notion of time is not relevant). This level of detail considers traffic speed, flow, and density) and their relationships according to the macroscopic traffic flow diagram. Cases of using macro simulation include:
- Binary integer programming (BIP)24: applied in estimating the temporal and spatial extent of delay caused by freeway accidents.
- Kinematic Shockwave propagation model22: the model leverages a physical traffic shockwave model, analysing different superposition situations of shockwaves. Also, this method has been compared to the car following model and found to be superior in performance (by almost ~20 times).
- Temporal Graph-Convolutional Network25 (a combination of gated recurrent units and graph-convolutional network) applied for real-time traffic forecasting.
The differential effects of determinants (traffic diversion requirement, crash injury type, number and type of vehicles involved in a crash, day of week and time of day, towing support requirement and damage to the infrastructure) on crash survival probabilities are found to vary considerably across the motorways26.
Both spatial and temporal aspects of traffic incidents were analysed using real-world spatiotemporal traffic sensor data on road networks27. Authors analysed abrupt and long-lasting propagation of the speed changes. As indicated by many researchers, calculating the consequences of an incident using micro-simulation models is extremely resource-intensive28–31, which made a number of experiments impossible.
As known, many transport simulation packages rely solely on the CPU usage (VISSIM, AIMSUN, SUMO, etc). However, recent advances in the development of parallel computing have made it possible to use GPUs in agent modelling and especially in traffic simulation. Studies of agent models32, and in particular transport micro-agents using GPUs, are already available31,33.
New opportunities which transport modelling on the GPUs opens up are significant and fundamental for further research in the area of traffic analysis. The use of GPUs opens a road for a field of incident response experiments involving optimisation based on micro-simulation.
Widely used traffic simulation software AIMSUN has been compared with FLAME GPU32. FLAME GPU shows significant speedup by orders of magnitude (10-100x). The largest simulation executed using both simulators, a one-hour simulation containing up to 512,000 vehicles and 1,575,936 detectors, showed a speedup of 43.8x for the GPU accelerated simulation. While one simulation is being performed on the CPU, up to ~50 simulations on the GPU can be done, so finding a solution strategy (by performing multiple simulations) for the incident becomes possible in an acceptable real-time.
Also, a simulation for the large-scale incident impact becomes possible. Thus, incident impact can be modelled both for small and large-scale environments. This approach has a potential to produce applicable incident response solutions in real-time within meaningful time constraints (during incident response planning). Use of the different tools for the task of traffic control (and not only traffic lights control) draws a complex optimisation task, which in combination with micro-simulation can produce more effective solutions than using only macro-models and simplistic control.
Stage 3. Incident management strategy search and evaluation
Context: Once an incident has been identified and verified, we need to find a strategy to reduce its impact on the traffic flow. In order to reduce the impact, different tools can be used: traffic light control methods, re-routing, speed limit adjustments, etc. Application of these tools can rely on different strategies and modes of operation.
Two approaches can be used for the task:
- Using computationally-efficient macro/meso models to evaluate different response plans; this option cannot fully grasp micro-behaviour of agents producing only shallow solutions to the task of impact reduction.
- Micro-modelling involving the use of GPUs, which allows to search for precise impact reduction strategy.
Depending on the predicted incident duration and severity, an actual placement of the incident in the road network, we can build an optimisation procedure to reduce the impact of an incident on the traffic flow.
Two-level traffic light control strategy (warning and stop lights) has been previously proposed for the prevention of incident-based traffic congestion with formulation of two different strategies for emergency and normal state traffic light control34. The value of this approach is that we can not only react to an actual incident but also develop pre-calculated congestion-prevention and response strategies for different kinds of incidents.
Incident management involves control of traffic network entities (e.g. traffic light, speed limits, re-routing) in order to reduce the impact of a traffic incident on the network traffic flow. Incident management can also be implemented in terms of dispatching and search for roaming strategy of response units29 with the aim to reduce their total travel time. Different approaches can be applied to find optimal traffic control strategy to optimise traffic flow including Deep Learning methods (such as Deep Reinforcement Learning (DRL)) and evolutionary optimisation35.
A special feature of reinforcement learning agents is the ability for online learning and transfer learning, which makes them suitable to constantly changing road conditions with ability to interchange learned information within the traffic control system. Recent work on the topic of robustness and uncertainty handling36, which relies on the use of callback-based RL framework (integrated with AIMSUN for testing traffic light control strategies).
Also, an open-source RL solution was presented for SUMO28. But the author states that RL methods were implemented only for traffic lights and not other types of control. In comparison to generic traffic light optimisation, for the task of incident management we will focus on the use of different tools for traffic flow control (not limited to traffic lights) during the event of an accident.
- The detection of traffic incidents, traffic incident severity and incident duration prediction (based on traffic flow parameters and incident log datasets)
- The spatial-temporal incident impact estimation (including impact map of adjacent transport elements and estimation of spatial-temporal lifecycle of disruption)
- A simulation-based modelling for response plan and incident clearance strategy. The solution would improve traffic centres decision making by automatic response plan recommendations.
The solution would improve traffic centres decision making by automatic response plan recommendations.
Please note …
Ongoing, this page will be a living record of this project. As it continues, matures, hits milestones, etc., we’ll add information, links, images, interviews and more. Watch this space!
- E. C. Sullivan, “New model for predicting freeway incidents and incident delays,” J. Transp. Eng., vol. 123, no. 4, pp. 267–275, 1997.
- Y. Chung, L. F. Walubita, and K. Choi, “Modeling accident duration and its mitigation strategies on South Korean freeway systems,” Transp. Res. Rec., vol. 2178, no. 1, pp. 49–57, 2010.
- K. Smith and B. L. Smith, “Forecasting the clearance time of freeway accidents,” 2002.
- D. Nam and F. Mannering, “An exploratory hazard-based analysis of highway incident duration,” Transp. Res. Part A Policy Pract., vol. 34, no. 2, pp. 85–102, 2000.
- A. T. Hojati, L. Ferreira, S. Washington, and P. Charles, “Hazard based models for freeway traffic incident duration,” Accid. Anal. Prev., vol. 52, pp. 171–181, 2013.
- Y. Zou, K. Henrickson, D. Lord, Y. Wang, and K. Xu, “Application of finite mixture models for analysing freeway incident clearance time,” Transp. A Transp. Sci., vol. 12, no. 2, pp. 99–115, 2016.
- R. Li, F. C. Pereira, and M. E. Ben-Akiva, “Competing risks mixture model for traffic incident duration prediction,” Accid. Anal. Prev., vol. 75, pp. 192–201, 2015.
- J. S. Yang, J. S. Lee, S. H. Park, D.-K. Kim, and K. S. Chon, “Accident Duration Distribution of Large-Vehicle Accidents on Korean Expressways,” J. East. Asia Soc. Transp. Stud., vol. 9, pp. 1855–1870, 2011.
- X. Mao, C. Yuan, J. Gan, and S. Zhang, “Risk Factors Affecting Traffic Accidents at Urban Weaving Sections: Evidence from China,” Int. J. Environ. Res. Public Health, vol. 16, no. 9, p. 1542, 2019.
- A.-S. Mihaita, Z. Liu, C. Cai, and M.-A. Rizoiu, “Arterial incident duration prediction using a bi-level framework of extreme gradient-tree boosting,” arXiv Prepr. arXiv1905.12254, 2019.
- K. Hamad, M. Khalil, and A. R. Alozi, “Predicting Freeway Incident Duration Using Machine Learning,” Int. J. Intell. Transp. Syst. Res., 2019.
- R. Prieto Curiel, H. González Ramírez, and S. R. Bishop, “A novel rare event approach to measure the randomness and concentration of road accidents,” PLoS One, vol. 13, no. 8, pp. 1–18, 2018.
- K. Ozbay and P. Kachroo, “Incident management in intelligent transportation systems,” 1999.
- W. Wang, H. Chen, and M. Bell, “A Study of the Characteristics of Traffic Incident Duration on Motorways,” 2002, pp. 1101–1108.
- K. Ozbay and N. Noyan, “Estimation of incident clearance times using Bayesian Networks approach,” Accid. Anal. Prev., vol. 38, no. 3, pp. 542–555, 2006.
- P. Barcellos, C. Bouvié, F. L. Escouto, and J. Scharcanski, “A novel video based system for detecting and counting vehicles at user-defined virtual loops,” Expert Syst. Appl., vol. 42, no. 4, pp. 1845–1856, 2015.
- S. Alkheder, M. Taamneh, and S. Taamneh, “Severity Prediction of Traffic Accident Using an Artificial Neural Network: Traffic Accident Severity Prediction Using Artificial Neural Network,” J. Forecast., vol. 36, 2016.
- X. Ma, C. Ding, S. Luan, Y. Wang, and Y. Wang, “Prioritizing Influential Factors for Freeway Incident Clearance Time Prediction Using the Gradient Boosting Decision Trees Method,” IEEE Trans. Intell. Transp. Syst., vol. 18, no. 9, pp. 2303–2310, 2017.
- J. de Oña, R. O. Mujalli, and F. J. Calvo, “Analysis of traffic accident injury severity on Spanish rural highways using Bayesian networks,” Accid. Anal. Prev., vol. 43, no. 1, pp. 402–411, 2011.
- F. Zong, H. Xu, and H. Zhang, “Prediction for traffic accident severity: comparing the Bayesian network and regression models,” Math. Probl. Eng., vol. 2013, 2013.
- H. Park, A. Haghani, and X. Zhang, “Interpretation of Bayesian neural networks for predicting the duration of detected incidents,” J. Intell. Transp. Syst., vol. 20, no. 4, pp. 385–400, 2016.
- C. Liu, S. Zhang, H. Wu, and Q. Fu, “A dynamic spatiotemporal analysis model for traffic incident influence prediction on urban road networks,” ISPRS Int. J. Geo-Information, vol. 6, no. 11, p. 362, 2017.
- M. Treiber, Traffic Flow Dynamics. 2013.
- Y. Chung and W. W. Recker, “A methodological approach for estimating temporal and spatial extent of delays caused by freeway accidents,” IEEE Trans. Intell. Transp. Syst., vol. 13, no. 3, pp. 1454–1461, 2012.
- L. Zhao, Y. Song, M. Deng, and H. Li, “Temporal graph convolutional network for urban traffic flow prediction method,” arXiv Prepr. arXiv1811.05320, 2018.
- H. Tajtehranifard, A. Bhaskar, M. M. Haque, and E. Chung, “Motorway crash duration and its determinants: do durations vary across motorways?,” J. Adv. Transp., vol. 50, no. 5, pp. 717–735, 2016.
- B. Pan, U. Demiryurek, C. Shahabi, and C. Gupta, “Forecasting spatiotemporal impact of traffic incidents on road networks,” in 2013 IEEE 13th International Conference on Data Mining, 2013, pp. 587–596.
- T. Azevedo, P. J. M. De Araújo, R. J. F. Rossetti, and A. P. C. Rocha, “JADE, TraSMAPI and SUMO: A tool-chain for simulating traffic light control,” arXiv Prepr. arXiv1601.08154, 2016.
- B. Huang and X. Pan, “GIS coupled with traffic simulation and optimization for incident response,” Comput. Environ. Urban Syst., vol. 31, pp. 116–132, 2007.
- K. Ozbay, W. Xiao, G. Jaiswal, B. Bartin, P. Kachroo, and M. Baykal-Gursoy, “Evaluation of incident management strategies and technologies using an integrated traffic/incident management simulation,” Int. J. Technol. Manag., 2009.
- P. Heywood, P. Richmond, and S. Maddock, “Road network simulation using FLAME GPU,” in European Conference on Parallel Processing, 2015, pp. 430–441.
- J. Xiao, P. Andelfinger, D. Eckhoff, W. Cai, and A. Knoll, “A Survey on Agent-Based Simulation using Hardware Accelerators,” ACM Comput. Surv., vol. 51, no. 6, p. 131, 2019.
- P. Heywood, S. Maddock, J. Casas, D. Garcia, M. Brackstone, and P. Richmond, “Data-parallel agent-based microscopic road network simulation using graphics processing units,” Simul. Model. Pract. Theory, vol. 83, pp. 188–200, 2018.
- L. Qi, M. Zhou, and W. Luan, “A Two-level Traffic Light Control Strategy for Preventing Incident-Based Urban Traffic Congestion,” IEEE Trans. Intell. Transp. Syst., vol. PP, pp. 1–12, 2016.
- M. Adriana Simona, L. Dupont, and M. Camargo, “Multi-objective traffic signal optimization using 3D mesoscopic simulation and evolutionary algorithms,” Simul. Model. Pract. Theory, vol. 86, 2018.
- F. Rodrigues and C. L. Azevedo, “Towards Robust Deep Reinforcement Learning for Traffic Signal Control: Demand Surges, Incidents and Sensor Failures,” arXiv Prepr. arXiv1904.08353, 2019.
- T. Salimans, J. Ho, X. Chen, S. Sidor, and I. Sutskever, “Evolution strategies as a scalable alternative to reinforcement learning,” arXiv Prepr. arXiv1703.03864, 2017.