Accident Anticipation via Temporal Occurrence Prediction
Traffic accident anticipation is a critical task for autonomous driving systems. We propose Temporal Occurrence Prediction (TOP), a novel framework that predicts when accidents will occur through temporal modeling. Our method enables real-time accident anticipation by modeling the temporal occurrence of potential collisions, providing timely alerts to enhance road safety.
The key insight is that modeling the temporal dynamics of accident occurrence allows for more accurate and actionable predictions in online scenarios. This leads to more reliable anticipation systems suitable for practical deployment.
We model accident anticipation through temporal occurrence prediction, capturing when accidents are likely to happen.
Our method performs real-time accident anticipation, suitable for practical deployment in autonomous driving systems.
The same architecture works across multiple datasets (CAP, DADA, Nexar) with consistent evaluation metrics.
Download the datasets and annotations for training and evaluation:
Our framework introduces comprehensive evaluation metrics for accident anticipation:
Average precision computed at different time windows (0.5s, 1.0s, 1.5s) before the accident. Measures classification accuracy at various anticipation horizons.
Mean AUC computed at FPR ≤ 10% across time windows. Evaluates true positive rate while maintaining low false alarm rate.
Measures how early the model detects accidents before they occur. Higher TTA indicates earlier and more actionable warnings.
We provide a reference implementation of our evaluation metrics. Key metrics include:
Download the metrics implementation:
Our method achieves state-of-the-art performance on multiple benchmarks:
Get started with TOP in three steps:
conda create -n top python=3.8.5 -y
conda activate top
pip install torch torchvision
pip install -U openmim
mim install mmengine==0.10.7
mim install mmcv==2.2.0git clone https://github.com/happytianhao/TOP.git
cd TOP
pip install -v -e .python tools/train.py configs/accident_anticipation/top/top_cap.pyIf you find this work useful, please consider citing:
@inproceedings{zhao2025accident,
title={Accident Anticipation via Temporal Occurrence Prediction},
author={Zhao, Tianhao and Zou, Yiyang and Mao, Zihao and Xiao, Peilun and Huang, Yulin and Yang, Hongda and Li, Yuxuan and Li, Qun and Wu, Guobin and Lin, Yutian},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2025}
}Check out our follow-up work on collision-anchored risk propagation:
RiskProp: Collision-Anchored Self-Supervised Risk Propagation for Early Accident Anticipation
Yiyang Zou, Tianhao Zhao, Peilun Xiao, Hongyu Jin, Longyu Qi, Yuxuan Li, Liyin Liang, Yifeng Qian, Chunbo Lai, Yutian Lin, Zhihui Li, Yu Wu
CVPR 2026 Highlight
@inproceedings{zou2026riskprop,
title={RiskProp: Collision-Anchored Self-Supervised Risk Propagation for Early Accident Anticipation},
author={Zou, Yiyang and Zhao, Tianhao and Xiao, Peilun and Jin, Hongyu and Qi, Longyu and Li, Yuxuan and Liang, Liyin and Qian, Yifeng and Lai, Chunbo and Lin, Yutian and Li, Zhihui and Wu, Yu},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2026}
}