Accident Anticipation via Temporal Occurrence Prediction

TL;DR

This paper introduces a novel accident anticipation method that predicts future accident scores at multiple time horizons using precise annotations, improving early warning accuracy and reducing false alarms in road safety applications.

Contribution

It shifts from binary risk scoring to multi-horizon accident score prediction using a Transformer-based model with precise supervision and a refined evaluation protocol.

Findings

Achieves higher recall at early time-to-accident intervals

Demonstrates improved TTA under practical false alarm constraints

Outperforms existing methods in accident anticipation accuracy

Abstract

Accident anticipation aims to predict potential collisions in an online manner, enabling timely alerts to enhance road safety. Existing methods typically predict frame-level risk scores as indicators of hazard. However, these approaches rely on ambiguous binary supervision (labeling all frames in accident videos as positive) despite the fact that risk varies continuously over time, leading to unreliable learning and false alarms. To address this, we propose a novel paradigm that shifts the prediction target from current-frame risk scoring to directly estimating accident scores at multiple future time steps (e.g., 0.1s-2.0s ahead), leveraging precisely annotated accident timestamps as supervision. Our method employs a snippet-level encoder to jointly model spatial and temporal dynamics, and a Transformer-based temporal decoder that predicts accident scores for all future horizons simultaneously using dedicated temporal queries. Furthermore, we introduce a refined evaluation protocol that reports Time-to-Accident (TTA) and recall (evaluated at multiple pre-accident intervals (0.5s, 1.0s, and 1.5s)) only when the false alarm rate (FAR) remains within an acceptable range, ensuring practical relevance. Experiments show that our method achieves superior performance in both recall and TTA under realistic FAR constraints.