Ifsttar PhD subject

French version
Detailed form :

Title : Modelling and stochastic simulation of autonomous vehicles' capacity to reduce driving risk

Main host Laboratory - Referent Advisor COSYS - PICS-L  -  SAINCT Remi      tél. : +33 130844049

Director of the main host Laboratory DUMONT Eric  -

PhD Speciality Statistiques, Automatique

Axis of the performance contract 1 - COP2017 - Efficient transport and safe travel

Main location Versailles-Satory

Doctoral affiliation UNIVERSITE PARIS - SACLAY

PhD school STIC - Sciences et technologies de l'Information et de la Communication

Planned PhD supervisor SAINT PIERRE Guillaume  -  Cerema  -  DTerSO - DALETT

Planned PhD co-supervisor GRUYER Dominique  -  Université Gustave Eiffel  -  COSYS - PICS-L

Planned financing Contrat doctoral  - Ifsttar

Abstract

General description :
In 2018, the Cerema, Ifsttar, LAB, Ceesar and Vedecom institute started the SURCA project (road users safety & automated driving), working for ONISR in France. They were to identify the interaction scenarios between the future autonomous vehicles (AVs) and other road users (non-autonomous vehicles, two-wheelers, pedestrians, etc.), with the idea of finding the conditions allowing AVs to perform as well as human drivers in normal driving conditions, and especially, better than them in an emergency, accident or near-accident situation. To reach social acceptance, AVs will have to prove their performance in handling these critical situations. However, a recent study [Kalra & Paddock 2016] showed we would need to collect the data of more than 100 million kilometers, or even 100 billions for some situations, to prove unambiguously the safety of autonomous cars. In spite of international effort, the AV industry is far from that, and it is necessary to explore other ways to evaluate the safety performances of AVs.
This thesis will use the recent evolution of simulation tools to model the behavior of a AV, and the knowledge of the most frequent accident scenarios from project Surca, to propose an new method to generate random risky scenarios, allowing to evaluate the efficiency of the AV. The ability of the simulation tool to reproduce human behavior and AV behavior allows the comparison of the performance of each group in terms of risk management in a critical situation. Thousands of random variations of a scenario, for different risky scenarios, will give the statistical significance needed for realistic conclusions.
The realism of a simulation depends on the chosen models and parameters. Specifically, it is crucial in this case to make sure the selected scenarios are representative (i.e. these situations actually occur in accident databases, cf project Surca), and interesting from a road safety perspective (using an abundant literature, for example Brenac & Clabaux 2010). The parameter distributions in each scenario will be chosen using the results of past naturalistic driving experiments (euroFOT, Udrive, ecoDriver, etc). Such parameters include driving speed depending on the infrastructure, acceleration/breaking profiles, reaction times, and countermeasures.
It will also be crucial to have a simplified but realistic model of the human driver, as well as the AV. The proximity of the Ifsttar management with the LaPEA (ex-LPC) and the Vedecom Institute will allow the candidate to acquire the knowledge necessary for this work. Knowledge of the ProSivic software by the Livic team will allow for a detailed modeling of the AV's behavior. Finally, the chosen stochastic model should guarantee the representativity of the scenarios and generate realistic situations. Methods like Monte Carlo and their stochastic versions, MCMC or Adaptive importance sampling (Robert & Casella 2004) could be used.
The results should on one hand give tangible elements on the increase of road safety using AVs (for example by comparing the number of avoided accidents between AVs and human drivers), but also draw attention on specific scenarios where this increase would be lower than expected.

State of the art :
The study of the first accident reports involving autonomous vehicles observed during the first large-scale experiments in the US is a major topic the the scientific and industrial community (Wang & Li 2019), but the number of observations, a few hundreds, is still too small to reassure the general public.
Simulation, and in particular the ProSivic software, is already a key tool in the conception and validation of driving assistance systems (by virtual prototyping), or in the study of connected vehicles (Bounini et al. 2014). Connecting real situations and virtual testing is known to be necessary for the study of accidents (Gechter et al. 2014).
The modeling of an AV is complicated by the fact that recent developments are private and confidential. However, some public tools exists giving the basing constitutive elements of an AV (Kato et al. 2015).

Expected results :
The first result will be methodological. The full modelling of an AV, combining different pieces of software (lane marking detection and object recognition for example) will be a major result by itself. The stochastic method for the generation of a number of realistic scenarios starting from a real situation (or a typology of observed accidents) will be a valuable result for the traffic simulation community.
The second type of results will be operational. The ability to generate a variety of interaction situations between users (AVs, other vehicles, pedestrians) in a reasonable computation time is extremely valuable for virtual prototyping in general.
The last type of results is scientific knowledge. The comparison, on thousands of realistic and potentially risky situations, between a human driven vehicle and an AV, will give a quantitative evaluation of the difference in terms of risk mitigation abilities. This "gain" (estimated by counting the situations leading to a shock for each type of vehicle) could be higher or lower depending on the type of scenario, indicating which situations are the most complex for the AV to deal with (though these could be very simple for a human).
In terms of non-scientific results, the interest of the ministry of interior, in particular the ONISR, is very high on these questions. Similarly, national car-making industries might be interested in the conclusions. The spinoffs will be on one hand a decision-support tool for leaders, and on the other hand a technological and strategic orientation for industrials.

Program of the thesis:
This thesis will include 6 steps:
1. State of the art of the accident scenarios that are the most critical to AVs: this work will mostly use the results of the Surca project with G. Saint Pierre, and all available literature (including reports for large-scale experiments in the US). This will allow to know the most likely interaction scenarios, and the aggravating factors that have the highest impact (weather conditions, blind spots, presence of vulnerable users, etc.). This will allow the definition of the situations to simulate, and give realistic distributions for the influencing parameters.
2. Study of stochastic simulation methods: the candidate will have to become familiar with Monte Carlo methods and their stochastic variants (including MCMC). Carefully choosing the distributions of the parameters influencing the scenario will allow, after many iterations, to get a distribution of the probability of accident for each type of vehicle and each type of scenario.
3. Learning to use the ProSivic software: this complex tool needs some time to become familiar with and a geographical relocalization in Versailles-Satory (after a first year in Toulouse).
4. Modeling of the AV system: once familiar with the software, the candidate will have to combine different available codes, build the non-existing ones, and create virtual environments for each scenarios.
5. Implementation of the complete test protocol, and simulation.
6. Analyze of the results, valorization and communication.
Steps 1 and 2 could be done during the first year, steps 3 and 4 will take most of the second year, the third one being used for the actual simulations and their valorization (steps 5 and 6).

Références :
Kalra, N., & Paddock, S. (2016). Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability? RAND Corporation.

Brenac T., Clabaux N., « Scénarios types d'accidents urbains n'impliquant pas de piétons et perspectives pour leur prévention. », INRETS, Département Mécanismes d'accidents, Edition : Bron : INRETS - 2010

Robert, C.P. and Casella, G. (2004) Monte Carlo Statistical Methods. Springer, New York.
Wang S, Li Z (2019) Exploring the mechanism of crashes with automated vehicles using statistical modeling approaches. PLoS ONE 14(3): e0214550.
BOUNINI F., GINGRAS D., LAPOINTE V., GRUYER D., "Real-time simulator of collaborative autonomous vehicles", IEEE Int. Conf. on Advances in Computing, Comm. and Informatics ICACCI, Greater Noida, India, pp. 723 - 729, September 2014.
Gechter, F., Dafflon, B., Gruer, P., & Koukam, A. (2014). Towards a Hybrid Real/Virtual Simulation of Autonomous Vehicles for Critical Scenarios.
S. Kato, E. Takeuchi, Y. Ishiguro, Y. Ninomiya, K. Takeda and T. Hamada, "An Open Approach to Autonomous Vehicles," in IEEE Micro, vol. 35, no. 6, pp. 60-68, Nov.-Dec. 2015.

Keywords : Autonomous vehicles, accidents, simulation, MCMC

Main host Laboratory - Referent Advisor	COSYS - PICS-L - SAINCT Remi tél. : +33 130844049
Director of the main host Laboratory	DUMONT Eric -
PhD Speciality	Statistiques, Automatique
Axis of the performance contract	1 - COP2017 - Efficient transport and safe travel
Main location	Versailles-Satory
Doctoral affiliation	UNIVERSITE PARIS - SACLAY
PhD school	STIC - Sciences et technologies de l'Information et de la Communication
Planned PhD supervisor	SAINT PIERRE Guillaume - Cerema - DTerSO - DALETT
Planned PhD co-supervisor	GRUYER Dominique - Université Gustave Eiffel - COSYS - PICS-L
Planned financing	Contrat doctoral - Ifsttar
Abstract General description : In 2018, the Cerema, Ifsttar, LAB, Ceesar and Vedecom institute started the SURCA project (road users safety & automated driving), working for ONISR in France. They were to identify the interaction scenarios between the future autonomous vehicles (AVs) and other road users (non-autonomous vehicles, two-wheelers, pedestrians, etc.), with the idea of finding the conditions allowing AVs to perform as well as human drivers in normal driving conditions, and especially, better than them in an emergency, accident or near-accident situation. To reach social acceptance, AVs will have to prove their performance in handling these critical situations. However, a recent study [Kalra & Paddock 2016] showed we would need to collect the data of more than 100 million kilometers, or even 100 billions for some situations, to prove unambiguously the safety of autonomous cars. In spite of international effort, the AV industry is far from that, and it is necessary to explore other ways to evaluate the safety performances of AVs. This thesis will use the recent evolution of simulation tools to model the behavior of a AV, and the knowledge of the most frequent accident scenarios from project Surca, to propose an new method to generate random risky scenarios, allowing to evaluate the efficiency of the AV. The ability of the simulation tool to reproduce human behavior and AV behavior allows the comparison of the performance of each group in terms of risk management in a critical situation. Thousands of random variations of a scenario, for different risky scenarios, will give the statistical significance needed for realistic conclusions. The realism of a simulation depends on the chosen models and parameters. Specifically, it is crucial in this case to make sure the selected scenarios are representative (i.e. these situations actually occur in accident databases, cf project Surca), and interesting from a road safety perspective (using an abundant literature, for example Brenac & Clabaux 2010). The parameter distributions in each scenario will be chosen using the results of past naturalistic driving experiments (euroFOT, Udrive, ecoDriver, etc). Such parameters include driving speed depending on the infrastructure, acceleration/breaking profiles, reaction times, and countermeasures. It will also be crucial to have a simplified but realistic model of the human driver, as well as the AV. The proximity of the Ifsttar management with the LaPEA (ex-LPC) and the Vedecom Institute will allow the candidate to acquire the knowledge necessary for this work. Knowledge of the ProSivic software by the Livic team will allow for a detailed modeling of the AV's behavior. Finally, the chosen stochastic model should guarantee the representativity of the scenarios and generate realistic situations. Methods like Monte Carlo and their stochastic versions, MCMC or Adaptive importance sampling (Robert & Casella 2004) could be used. The results should on one hand give tangible elements on the increase of road safety using AVs (for example by comparing the number of avoided accidents between AVs and human drivers), but also draw attention on specific scenarios where this increase would be lower than expected. State of the art : The study of the first accident reports involving autonomous vehicles observed during the first large-scale experiments in the US is a major topic the the scientific and industrial community (Wang & Li 2019), but the number of observations, a few hundreds, is still too small to reassure the general public. Simulation, and in particular the ProSivic software, is already a key tool in the conception and validation of driving assistance systems (by virtual prototyping), or in the study of connected vehicles (Bounini et al. 2014). Connecting real situations and virtual testing is known to be necessary for the study of accidents (Gechter et al. 2014). The modeling of an AV is complicated by the fact that recent developments are private and confidential. However, some public tools exists giving the basing constitutive elements of an AV (Kato et al. 2015). Expected results : The first result will be methodological. The full modelling of an AV, combining different pieces of software (lane marking detection and object recognition for example) will be a major result by itself. The stochastic method for the generation of a number of realistic scenarios starting from a real situation (or a typology of observed accidents) will be a valuable result for the traffic simulation community. The second type of results will be operational. The ability to generate a variety of interaction situations between users (AVs, other vehicles, pedestrians) in a reasonable computation time is extremely valuable for virtual prototyping in general. The last type of results is scientific knowledge. The comparison, on thousands of realistic and potentially risky situations, between a human driven vehicle and an AV, will give a quantitative evaluation of the difference in terms of risk mitigation abilities. This "gain" (estimated by counting the situations leading to a shock for each type of vehicle) could be higher or lower depending on the type of scenario, indicating which situations are the most complex for the AV to deal with (though these could be very simple for a human). In terms of non-scientific results, the interest of the ministry of interior, in particular the ONISR, is very high on these questions. Similarly, national car-making industries might be interested in the conclusions. The spinoffs will be on one hand a decision-support tool for leaders, and on the other hand a technological and strategic orientation for industrials. Program of the thesis: This thesis will include 6 steps: 1. State of the art of the accident scenarios that are the most critical to AVs: this work will mostly use the results of the Surca project with G. Saint Pierre, and all available literature (including reports for large-scale experiments in the US). This will allow to know the most likely interaction scenarios, and the aggravating factors that have the highest impact (weather conditions, blind spots, presence of vulnerable users, etc.). This will allow the definition of the situations to simulate, and give realistic distributions for the influencing parameters. 2. Study of stochastic simulation methods: the candidate will have to become familiar with Monte Carlo methods and their stochastic variants (including MCMC). Carefully choosing the distributions of the parameters influencing the scenario will allow, after many iterations, to get a distribution of the probability of accident for each type of vehicle and each type of scenario. 3. Learning to use the ProSivic software: this complex tool needs some time to become familiar with and a geographical relocalization in Versailles-Satory (after a first year in Toulouse). 4. Modeling of the AV system: once familiar with the software, the candidate will have to combine different available codes, build the non-existing ones, and create virtual environments for each scenarios. 5. Implementation of the complete test protocol, and simulation. 6. Analyze of the results, valorization and communication. Steps 1 and 2 could be done during the first year, steps 3 and 4 will take most of the second year, the third one being used for the actual simulations and their valorization (steps 5 and 6). Références : Kalra, N., & Paddock, S. (2016). Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability? RAND Corporation. Brenac T., Clabaux N., « Scénarios types d'accidents urbains n'impliquant pas de piétons et perspectives pour leur prévention. », INRETS, Département Mécanismes d'accidents, Edition : Bron : INRETS - 2010 Robert, C.P. and Casella, G. (2004) Monte Carlo Statistical Methods. Springer, New York. Wang S, Li Z (2019) Exploring the mechanism of crashes with automated vehicles using statistical modeling approaches. PLoS ONE 14(3): e0214550. BOUNINI F., GINGRAS D., LAPOINTE V., GRUYER D., "Real-time simulator of collaborative autonomous vehicles", IEEE Int. Conf. on Advances in Computing, Comm. and Informatics ICACCI, Greater Noida, India, pp. 723 - 729, September 2014. Gechter, F., Dafflon, B., Gruer, P., & Koukam, A. (2014). Towards a Hybrid Real/Virtual Simulation of Autonomous Vehicles for Critical Scenarios. S. Kato, E. Takeuchi, Y. Ishiguro, Y. Ninomiya, K. Takeda and T. Hamada, "An Open Approach to Autonomous Vehicles," in IEEE Micro, vol. 35, no. 6, pp. 60-68, Nov.-Dec. 2015.
Keywords :	Autonomous vehicles, accidents, simulation, MCMC

List of topics

Applications closed