Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Increasing Driving Safety : Perception & Assessment of Collision Risks with other Road Users Christian Laugier To cite this version: Christian Laugier. Increasing Driving Safety : Perception & Assessment of Collision Risks with other Road Users. Workshop ”Safety of Autonomous Vehicles”, Final scientific event of the French Tornado project, Nov 2020, Paris & on-line event, France. ฀hal-03147801฀ HAL Id: hal-03147801 https://hal.inria.fr/hal-03147801 Submitted on 20 Feb 2021 HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Increasing Driving Safety Perception & Assessment of Collision Risks with other Road Users Dr. Christian LAUGIER Research Director at Inria (christian.laugier@inria.fr) Invited Talk Workshop “Safety of Autonomous Vehicles” Final Scientific event of the French Tornado R&D project Paris, On-line event, November 5th 2020 C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 © Laugier & Inria. All rights reserved 1 Increase Driving Safety Perception & Assessment of Collision Risks with other Road Users C. Laugier, Research Director at Inria Workshop « Safety of Autonomous Vehicles », Tornado project, Nov 5th 2020 • Technological breakthrough & Numerous AV experimentations in real traffic conditions Tesla Autopilot L2 (ADAS) • 32 millions km covered since 2009 • 1st US self-driving taxi service L3 (Phoenix, since Dec 2018) Robot taxi Autonomous Shuttle service • Millions of miles driven last decade… but SAFETY is still not fully guaranteed ! => Several benign or fatal accidents involving AVs (Perception failure & Takeover difficult) Tesla Autopilot level 2 (May 2016) Self-driving Uber L3 (March 2018) C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 => Perception & Decision-making technologies have still to be improved for mixt traffic !!! 2 © Laugier & Inria. All rights reserved Perception & Situation Awareness Challenges ADAS & Autonomous Driving Embedded Perception & Safe driving Road Safety Campaign, France 2014 Dealing with unexpected events => Avoiding upcoming collisions with “something” Situation Awareness & Decision-making => Safe intentional navigation (using semantics) Main difficulties  Dynamic & Open Environments, Incompleteness & Uncertainty, Sensors limitations, Real-time + Validation  Mixed traffic (Human in the loop) => Human Aware Decision-making process Taking into account Interactions + Behaviors + Social & Traffic rules C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 © Laugier & Inria. All rights reserved 3 1st Paradigm: Embedded Bayesian Perception Sensors Fusion => Mapping & Detection Velodyne 3D lidar Characterization of the local Safe Navigable Space & Collision Risk cameras IBEO lidars Embedded Multi-Sensors Perception  Continuous monitoring of the dynamic environment Dynamic scene interpretation => Using Context & Semantics  Exploiting the Dynamic information for a better understanding of the scene !!!!  Reasoning about Uncertainty & Time window => Past & Future predicted events  – Confiance numérique: le cas du véhicule autonome C. LAUGIER AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 Bayesian Sensors Fusion + Scene interpretation using Contextual & Semantic information © Laugier & Inria. All rights reserved 4 2nd Paradigm: Collision Risk Assessment => Avoiding Pending & Future Collisions Human-Aware Situation Assessment Complex dynamic situation Risk-Based Decision-making => Safest maneuver to execute Alarm / Control Several PhD theses  Predict environment changes on a given “time horizon t+d” => Using History & Motion models  Estimate the Probabilistic Risk of Collision at t+d (d = a few seconds ahead)  Make Driving Decisions by taking into account the Predicted behavior of all surrounding traffic participants (cars, cycles, pedestrians…) & Social / Traffic rules (traffic participants interactions) Real world Static Risk /Alarm Dynamic • Risk Location • Collision Probability • TTC Moving Dummy Observed moving Car Camera view (in ego vehicle) 1s before the crash No risk (White car) =>safe motion direction Video: Collision Risk Assessment • Yellow => time to collision: 3s • Orange => time to collision: 2s • Red => time to collision: 1s High risk (Pedestrian) C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 © Laugier & Inria. All rights reserved 5 3rd Paradigm: Models improvements using Machine Learning  Perception level: Construct “Semantic Grids” using Bayesian Perception & DL RGB images (for semantic segmentation) 3D Point clouds (for Dynamic Occupancy Grids) Semantic Grids  Prediction & Decision-making level: Learn driving skills for Autonomous Driving o 1st Step: Modeling Driver Behaviors using IRL Ego vehicle Front cam  White vehicle => Ego-vehicle (ground-truth) Ego vehicle Back cam  Red box => Plan induced (Predicted trajectory)  Yellow boxes => Detected obstacles (CMCDOT) o 2nd Step:  Predict behaviors of surrounding vehicles (using Perception & learned Behavior models)  Make “safe & consistent” Driving Decisions for Ego Vehicle  Open questions: Training step (Available Datasets limited), Real-time processing (difficult), Classification Errors (often not explainable), Domain adaptation (e.g. changing weather conditions) C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 © Laugier & Inria. All rights reserved 6 Concluding remarks & Discussion  Increasing impact of AI + Real-time data processing capacity + Increased sensor performance + New Models & Embedded algorithms + Multiplication of tests in real conditions => The unmanned car is gradually becoming a technological reality  Safety is not yet fully guaranteed ! o Current Perception & Scene Understanding algorithms are not robust enough for complex & highly dynamic environments o Need to take better account of Interactions with other road users (using also AI approaches) o Need to develop Validation & Certification Tools and Methodologies => Realistic simulators, Real- world testing protocols, Formal methods (e.g. Enable-S3 EU project & future French project Prissma)  User confidence & Acceptance by the human society will be decisive to allow a real deployment (e.g. “cohabitation” with other users such as pedestrians, bicycles, scooters ...) o Autonomous vehicles a priori safer than cars driven by humans (inattention)…. but 0 tolerance in the event of a fatal accident involving an autonomous vehicle ! Ethics & Responsibility issues must also be taking into account before any deployment o C. LAUGIER – Confiance numérique: le cas du véhicule autonome AG IRT-Nanoelec – Table ronde “Une filière pour une société résiliante”, Minatec Grenoble, Sept 8th 2020 © Laugier & Inria. All rights reserved 7