MASTTA: A Multi-modal Architecture with Spatio-Temporal-Text Adaptation for Video-based Traffic Accident Anticipation

This is an official repo implementation of A Multi-modal Architecture with Spatio-Temporal-Text Adaptation for Video-based Traffic Accident Anticipation, IEEE Transactions on Circuits and Systems for Video Technology

The code and pre-trained weights will be released soon...

Architecture

Visualisation of Accident Events

Results

Performance comparison of the achievable longest mTTA and the corresponding mAP of state-of-the-art models on the DAD and CCD datasets.
The best balanced results are in bold.

DAD Dataset Method Source mAP (%) mTTA (s) DSA ACCV16 48.4 1.34 adaLEA CVPR18 52.3 3.43 GCRNN ACMMM20 53.7 3.53 FA ICPR21 49.8 3.76 DRIVE ICCV21 62.8 2.78 L-RA CVPR21 49.1 3.04 DSTA TITS22 56.1 3.66 GSC TIV23 60.4 2.55 DAA-GNN PR23 70.6 1.59 - ICIP24 61.1 4.06 MASTTA (Ours) - 70.2 3.96

CCD Dataset Method Source mAP (%) mTTA (s) DSA ACCV16 99.6 4.87 GCRNN ACMMM20 99.5 4.74 DSTA TITS22 99.6 4.52 - ICIP24 99.3 4.97 MASTTA (Ours) - 99.9 4.95

---

Performance comparison of the achievable highest mAP and the corresponding mTTA of state-of-the-art models on the DAD dataset.
The best balanced results are in bold.

Method	Source	mAP (%)	mTTA (s)	TTA₈₀R
GCRNN	ACMMM20	72.23	1.33	1.56
L-RAI	CVPR21	51.40	3.01	-
MViTv2	CVPR21	64.49	-	-
DSTA	TITS22	72.34	1.50	1.81
VideoSwin	CVPR22	65.45	-	-
UniFormer V2	ICCV23	65.24	-	-
DAA-GNN	PR23	75.20	1.47	-
-	ICIP24	75.61	3.31	-
MASTTA (Ours)	-	80.81	3.32	3.51

Citation

If you find our work useful, plase use the following BiBtex for citation:

@ARTICLE{10933925,
  author={Patera, Patrik and Chen, Yie-Tarng and Fang, Wen-Hsien},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={A Multi-modal Architecture with Spatio-Temporal-Text Adaptation for Video-based Traffic Accident Anticipation}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Visualization;Accidents;Adaptation models;Feature extraction;Attention mechanisms;Accuracy;Tuning;Data models;Computational modeling;Vehicle dynamics;accident anticipation;multi-modal architecture;parameter-efficient transfer learning;image-to-video adaptation},
  doi={10.1109/TCSVT.2025.3552895}
}