ROAD: The Road Event Awareness Dataset for Autonomous Driving

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Human driving involves holistic understanding of dynamic road events and their evolution.
• Current autonomous vehicles (AVs) lack human-level situational awareness and decision-making capabilities.
• Bridging this gap requires AVs to perceive and interpret complex road events.

Purpose of the Study:

• Introduce the ROad event Awareness Dataset (ROAD), the first dataset of its kind for autonomous driving.
• Enable testing and benchmarking of AVs' ability to detect and understand road events.
• Facilitate research in areas like complex activity detection, future event anticipation, and continual learning for AVs.

Main Methods:

• Utilized videos from the Oxford RobotCar Dataset, annotated with bounding boxes for road events.
• Defined road events as triplets: active agent, action(s), and scene location.
• Proposed 3D-RetinaNet, an incremental algorithm for online road event awareness, as a baseline.
• Benchmarked performance using 3D-RetinaNet, Slowfast, YOLOv5, and ICCV2021 ROAD challenge winners.

Main Results:

• The ROAD dataset provides a novel benchmark for evaluating road event detection in AVs.
• Performance benchmarks highlight the challenges in achieving human-level situation awareness for AVs.
• 3D-RetinaNet demonstrates a promising baseline for online road event awareness.

Conclusions:

• The ROAD dataset is crucial for advancing situational awareness in autonomous driving.
• Further research using ROAD can lead to more sophisticated AV perception and decision-making.
• The dataset and baseline algorithm are publicly available to foster academic investigation.