ROAD: ROAD EVENT AWARENESS DATASET FOR INTELLIGENT VEHICLE PERCEPTION

Abstract

Humans drive holistically, which involves understanding dynamic road events and how they evolve over time, among other things. These characteristics may help self-driving vehicles make more human-like decisions and have higher situational awareness. We provide the first-ever ROad event Awareness Dataset (ROAD) for Autonomous Driving to achieve this. Evaluation of an autonomous vehicle's ability to identify road events—triplets consisting of an active agent, the action or acts it does, and the scenes that correspond to those actions—is the goal of ROAD. ROAD consists of films originally extracted from the Oxford RobotCar Dataset, annotated with bounding boxes that show the position of each road event in the image plane. As a starting point for various detection tasks, we suggest 3D-RetinaNet, a unique incremental approach for online road event awareness.
In order to illustrate the challenges situation awareness in autonomous driving confronts, we also provide the results of the ICCV2021 ROAD challenge winners as well as the performance of the Slowfast and YOLOv5 detectors on the ROAD tasks. The purpose of ROAD is to allow academics to investigate intriguing problems such as advanced (road) activity detection, future event prediction, and continuous learning.