ROAR: Robust accident recognition and anticipation for autonomous driving

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Autonomous vehicle (AV) safety relies on accurate accident anticipation.
• Existing methods fail under real-world conditions like sensor failures, noise, and data imperfections.
• Variability in driver behavior and accident rates across vehicle types is not adequately addressed.

Purpose of the Study:

• Introduce ROAR, a novel approach for robust accident detection and prediction in AVs.
• Address limitations of existing methods in handling imperfect data and diverse traffic scenarios.
• Enhance the reliability and accuracy of AV safety systems.

Main Methods:

• Utilize Discrete Wavelet Transform (DWT) for feature extraction from noisy/incomplete data.
• Employ a self-adaptive object-aware module to focus on high-risk vehicles and model spatial-temporal relationships.
• Implement dynamic focal loss to address class imbalance between accident and non-accident events.

Main Results:

• ROAR consistently outperforms existing baselines on the Dashcam Accident Dataset (DAD), Car Crash Dataset (CCD), and AnAn Accident Detection (A3D).
• Achieved superior performance in Average Precision (AP) and mean Time-to-Accident (mTTA).
• Demonstrated robustness in handling sensor degradation, environmental noise, and imbalanced data.

Conclusions:

• ROAR offers a significant advancement in accident anticipation for autonomous vehicles.
• The model's robustness makes it suitable for complex and unpredictable real-world traffic environments.
• Provides a promising solution for enhancing the safety and reliability of autonomous driving systems.