Learning to Annotate Delayed and False AEB Events: A Practical System for Extreme Class Imbalance and Asymmetric Label Noise

Problem
The paper addresses the challenge of accurately annotating rare but critical delayed and false Autonomous Emergency Braking (AEB) events, which constitute less than 5% of daily AEB triggers. Manual annotation is prohibitively expensive at scale, and the authors highlight that existing literature lacks automated solutions for this specific problem. This work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The authors propose an automated AEB annotation framework that incorporates two primary innovations to tackle the identified challenges. First, they implement a specific data augmentation strategy that synthesizes realistic samples by manipulating focal target attributes, transplanting ego-vehicle dynamics, and masking non-focal agents. This approach aims to enhance the representation of delayed and false triggers in the training dataset. Second, they introduce a noise suppression mechanism that employs stable hardness estimation and probe-guided adaptive thresholds to clean mislabeled true trigger samples. The framework is designed as a full-stack architecture, enabling efficient identification of critical delayed and false triggers from thousands of daily AEB events.

Results
The proposed system demonstrates significant performance improvements, achieving an 80% increase in recall for delayed and false triggers compared to baseline methods. Additionally, it results in a 50% reduction in manual annotation workload. These results indicate a substantial enhancement in the annotation process’s efficiency and accuracy, which is critical for optimizing AEB systems in real-world applications.

Limitations
The authors acknowledge that the framework’s performance may be influenced by the quality of the initial dataset and the effectiveness of the data augmentation techniques. They do not address potential scalability issues when deployed in diverse driving environments or the need for continual adaptation to evolving driving conditions. Furthermore, the reliance on synthetic data may introduce biases that are not present in real-world scenarios.

Why it matters
This work has significant implications for the development of robust AEB systems, as it establishes a necessary data foundation for ongoing optimization efforts. By automating the annotation of critical events, the framework not only reduces the manual workload but also facilitates continuous self-improvement through the accumulation of high-quality annotations. This advancement is crucial for enhancing the safety and reliability of autonomous driving technologies, as highlighted in the context of AEB systems. The findings and methodologies presented in this paper are relevant for future research in automated annotation and machine learning applications in safety-critical systems, as published in arXiv cs.LG.