Cooperative Autonomous Driving and Interaction with Vulnerable Road Users

Transcription

1 9th Workshop on PPNIV Keynote Cooperative Autonomous Driving and Interaction with Vulnerable Road Users Miguel Ángel Sotelo Full Professor University of Alcalá (UAH) SPAIN 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept. 2017

2 Content Motivation Autonomous Cooperative Driving GCDC Results Interaction with VRUs Conclusions and Future Work 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

3 Motivation Despite the great development in the past years, there are still some major limitations in Autonomous Driving: Legal framework Navigation Reliability Efficiency Social Acceptance 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

4 Limitations of Autonomous Vehicles Navigation: - Enriched maps are needed (2 Gb/Km). - International Consortia: BMW, Daimler, Audi (HERE). - Online data acquisition and map building. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

5 Limitations of Autonomous Vehicles Reliability: - Improvement in sensorial capabilities (adverse weather). - Development of Cooperative Systems. Efficiency: - Human-like decision making and maneuvering. - Emulation of human driving by means of prediction of intentions of other traffic agents, such as pedestrians and other vehicles. - There is a need for enhanced cooperation and interaction capabilities. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

6 Content Motivation Autonomous Cooperative Driving GCDC Results Interaction with VRUs Conclusions and Future Work 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

7 Goal: Autonomous Cooperative Driving - Increase reliability and efficiency of autonomous vehicles. Techniques: - Cooperation with other vehicles (autonomously or manually driven) and with the infrastructure. - Cooperation with VRUs (Vulnerable Road Users) by prediction their intentions and trajectories. Limitations: - Strong dependency on penetration rate. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

8 Initiatives: Autonomous Cooperative Driving - European Commission: funding of research projects on Cooperative Systems and Autonomous Driving (FP7 and H2020). - Grand Cooperative Driving Challenge (GCDC): International Competition on Autonomous Cooperative Driving in Helmond (The Netherlands) in 2011 and GCDC 2016 (three tests): - Platooning + Merging. - Management of T-intersections. - Management of emergency vehicles. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

9 Autonomous Cooperative Driving GCDC 2016: Platooning + Merging 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

10 Autonomous Cooperative Driving GCDC 2016: Management of T intersections 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

11 Autonomous Cooperative Driving GCDC 2016: Management of emergency vehicles 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

12 Autonomous Cooperative Driving DRIVERTIVE General Architecture 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

13 Autonomous Cooperative Driving Data Fusion - Localization Example 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

14 Autonomous Cooperative Driving Communications System (ITS G5 V2V standard) Messages types CAM (Cooperative Awareness Message): position, geometry and vehicles dynamics. DENM (Decentralized Environmental Notification Message): asynchronous messages from infrastructure o from other vehicles (e.g. emergency vehicles approaching). iclcm (igame Cooperative Lane Change Message): messages for interaction protocol in different scenarios. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

15 Autonomous Cooperative Driving Scenario 1: Platooning + Merging 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

16 Autonomous Cooperative Driving Scenario 1: Platooning + Merging Behavior on the left lane is different from that on the right lane 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

17 Autonomous Cooperative Driving Scenario 2: Management of T-intersections 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

18 Autonomous Cooperative Driving Scenario 2: T-Intersections A safety distance must be kept at all times w.r.t the preceding vehicle 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

19 Autonomous Cooperative Driving Analysis of the communication channel (CCDF Complementary Cumulative Distribution Function) 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

20 Autonomous Cooperative Driving Mean and Variance of UD (Car) 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

21 Autonomous Cooperative Driving Mean and Variance of UD (Truck) 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

22 Autonomous Cooperative Driving Analysis of the communication channel - For some vehicles, the probability of large delays is significant (>10%). - The UD degrades with distance. - Occlusions have a strong effect on delays: - Trucks are less occluded given that their antennas are located at a height of 3 meters. - Other findings: DCC in a highly congested channel is making some of the vehicles get stuck in Restrictive state and are not able to regularly access the channel. - CAM and DENM in GCDC at 25 Hz. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

23 Autonomous Cooperative Driving Effect of UD on Emergency Braking during CACC - There is a probability between of collision with the leading vehicle is only communications are used for CACC in a fleet of more than 4 vehicles. - The channel load is responsible for low reliability. - An additional sensor is needed (radar). 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

24 Content Motivation Autonomous Cooperative Driving GCDC Results Interaction with VRUs Conclusions and Future Work 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

25 Autonomous Cooperative Driving DRIVERTIVE University of Alcalá s team - Autonomous Cooperative Vehicle (Velodyne, Radar, 3D Vision, Laser, DGPS, CANBus, Communications, fully automated) 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

26 GCDC Results DRIVERTIVE at GCDC th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

27 GCDC Results DRIVERTIVE Winner of the Prize to the Best Team with Full Automation in GCDC th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

28 GCDC Results GCDC th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

29 Content Motivation Autonomous Cooperative Driving GCDC Results Interaction with VRUs Conclusions and Future Work 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

30 Motivation 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

31 Motivation Pedestrian Path Prediction in the Automotive: Further improvement in state-of-the-art ADAS by means of action classification Walking, Stopping, Starting, Bending-in Improvement Strong of accuracy gains are in expected cm: in the performance and reliability of active Difference pedestrian between protection effective systems and non-effective intervention in emergency braking systems Initiation of emergency braking 0.16 s in advance can potentially reduce severity of accidents injuries by 50% Early recognition of pedestrians stopping actions can provide more accurate last-second active interventions 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

32 Proposed Approach Global Scheme Off-line Motion Capture System Probabilistic Training Knowledge of Pedestrians dynamics On-line Stereo Cameras Geometric processing 3D Pedestrian Pose Estimation Transformation to latent space + prediction Lateral predicted position Recovery of 3D pose and position 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

33 Pedestrian Pose Measurement Pedestrian Skeleton considered in this research 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

34 Pedestrian Pose Measurement Method for Joints Extraction - Example 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

35 Pedestrian Pose Measurement Method for Joints Extraction Results 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

36 Pedestrian Pose Measurement Body parts detection using Deep Learning 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

37 General Method - Overview 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

38 Activity Recognition - Example 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

39 Experimental Results Detection Delay - Summary 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

40 Experimental Results Probabilistic Action Classification 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

41 Experimental Results Probabilistic Action Classification 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

42 Experimental Results Video sequence showing prediction results 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

43 Intelligent Interface with VRUs GRAIL GReen Assistant Interfacing Light 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

44 Intelligent Interface with VRUs GRAIL GReen Assistant Interfacing Light 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

45 Content Motivation Autonomous Cooperative Driving GCDC Results Interaction with VRUs Conclusions and Future Work 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept

46 Conclusions and Future Work Conclusions - Autonomous Cooperative Systems will pave the way to the massive and robust deployment of self-driving cars. - The V2V communication link is still a weakness that needs further attention from the scientific community in order to provide real-time and robust communication capability among large fleets of vehicles. - Anticipating the intentions of other traffic participants, such as VRUs and vehicles, is essential for mimicking human drivers behavior. Future Work - Enhancement of V2V communication channel for large fleets of vehicles (antenna placement, frequency of data, etc.). - Context-based action prediction using Probabilistic Graphical Models (Bayesian Networks) is under development for VRUs and vehicles. - Gaze direction, group behavior. 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept. 2017

47 9th Workshop on PPNIV Keynote Cooperative Autonomous Driving and Interaction with Vulnerable Road Users Thanks for your kind attention! Miguel Ángel Sotelo Full Professor University of Alcalá (UAH) SPAIN 9 th Workshop on Planning, Perception, and Navigation for Intelligent Vehicles. Vancouver, Canada, 24 th Sept. 2017