Research map / Label-efficient and robust learning / Recognition and long-tail

Progressive Risk Estimation for Accident Anticipation

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Reformulates accident anticipation as continuous, monotonically increasing risk estimation with progress regression and pairwise preference ranking on a fine-tuned VideoMAE, and proposes a Separation Score for global risk-curve evaluation.

Problem

Accident anticipation from dashcam video is usually framed as binary classification of whether a crash will occur, ignoring when it will occur and producing false alarms during normal driving; standard metrics (AUC, mTTA, mAP) evaluate only local windows.

Approach

Model accident risk as an exponentially shaped, monotonically increasing function of normalized time-to-accident within the critical zone, trained with BCE, a smooth L1 progress-regression loss, and a margin ranking loss enforcing higher risk closer to the accident; add a Separation Score that compares mean pre-onset and post-onset risk.

Key results

• On CAP, PRE-ACT reaches mAUC 0.481 vs 0.429 for TOP and mTTA 1.125 s vs 0.864 s (about 0.25 s earlier).
• On DADA, PRE-ACT reaches mAUC 0.440 vs 0.332 for TOP, with mTTA improved by 0.18 s.
• On Nexar, PRE-ACT achieves mAP 0.895, exceeding the challenge winner (0.885) by about 1 point.

Surprising findings

• The BCE-only baseline alone is already comparable to the prior SOTA TOP in mAUC (0.419 vs 0.429) while being much earlier in mTTA.
• Performance drops substantially at strict FPR<=0.01 (CAP mAUC 0.210), and cross-dataset transfer from CAP to Nexar degrades mAUC by about 0.15.

Limitations

• Studies accident anticipation only as video monitoring, ignoring driver state, vehicle dynamics and sensor reliability.
• Cross-domain generalization from CAP to Nexar is weak due to domain gap.
• Failures under poor visibility such as snow or strong sunlight; low-FPR performance drops substantially.

Experimental setup

Datasets: CAP (MM-AU) DADA (MM-AU) Nexar Kinetics-400

Benchmarks: MM-AU CAP MM-AU DADA Nexar

Metrics: AUC at 0.0/0.5/1.0/1.5s horizons (FPR<=0.1) mAUC mTTA mAP Separation Score

Baselines: CAP DRIVE DSTA GSC TOP Nexar challenge winner Nexar challenge runner-up

Builds on: VideoMAE TOP BCE classification

Closest papers by meaning

• Accident Anticipation via Temporal Occurrence Prediction (NeurIPS 2025 · similarity 0.92)
• Harnessing Event Sensory Data for Error Pattern Prediction in Vehicles: A Language Model Approach (AAAI 2025 · similarity 0.84)
• SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.83)
• Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.82)
• Monitoring of Perception Systems: Deterministic, Probabilistic, and Learning-Based Fault Detection and Identification (Abstract Reprint) (AAAI 2024 · similarity 0.82)
• Dual-View Interaction-Aware Lane Change Prediction for Autonomous Driving (AAAI 2025 · similarity 0.82)