Detecting Hazardous Events: A Framework for Automated Vehicle Safety Systems

Abstract

The driving domain is inherently dangerous. To develop connected and automated vehicles that can detect potential sources of harm, we must clearly define these hazardous events and metrics to detect them. The majority of driving scenarios we face do not materialise harm, but we often face potentially hazardous near-miss scenarios. Potential harm is difficult to quantify when harm is not materialised; thus, few metrics detect these scenarios in the absence of collision and even fewer datasets label non-collision-based hazardous events. This study focuses on detecting near-miss scenarios due to other actors since human error is the primary source of harm. We first provide a concise overview of current event-specific metrics. We then propose an event-agnostic detection framework that exploits vehicle kinematics to detect evasive manoeuvres early and dynamically calculate minimum safe distances. Given inconsistent dataset labelling methods and collision-focused events, we provide a preliminary study to demonstrate an eventagnostic and configurable dataset annotation technique to label hazardous events, even when harm is not materialised. We show promising results detecting hazardous scenes on a labelled simulation benchmark, GTACrash.

Introspection of DNN-Based Perception Functions in Automated Driving Systems: State-of-the-Art and Open Research Challenges

“…However, there might be different levels or types of where, where some might not require a safety response or be mitigated by other measures. Additionally, in certain situations, perception quality is of paramount importance for other safety-relate mechanisms, particularly in the event of a potential accident, hazardous situation [121], [122] or an irregular driver behaviour [123]. These scenarios warrant further examination, which can be achieved through introspection.…”

Review of Graph-Based Hazardous Event Detection Methods for Autonomous Driving Systems

“…To detect hazardous events, AV systems must process a vast amount of perception data (e.g., object detection, classification and localisation), which can be noisy, uncertain or incomplete, in order to identify potential events that may materialise harm. However, achieving early and robust hazardous event detection remains challenging due to an unlimited variety of edge cases, unseen environments, and the driving domain's overall complexity [4]- [7]. Aggravated by the fact that it is impossible to define all hazardous events manually, a scalable approach is vital to adapt in an ever-changing domain.…”

Introspection of 2D Object Detection using Processed Neural Activation Patterns in Automated Driving Systems

“…Deep neural networks (DNN) are becoming widely accepted for their improved object detection performance compared to traditional computer vision techniques. Nevertheless, their reliability and stability in safety-critical situations such as those encountered in ADS remain to be a challenge for system developers [1,17,26]. Even the best perception systems are not bulletproof; perception errors cannot be eliminated despite the rapid advancements in DNNs.…”