Traffic accidents are a leading cause of fatalities and injuries across the globe. Therefore, the ability to anticipate hazardous situations in advance is essential. Automated accident anticipation enables timely intervention through driver alerts and collision avoidance maneuvers, forming a key component of advanced driver assistance systems. In autonomous driving, such predictive capabilities support proactive safety behaviors, such as initiating defensive driving and human takeover when required. Using dash-cam video as input offers a cost-effective solution, but it is challenging due to the complexity of real-world driving scenes. Accident anticipation systems need to operate in real-time. However, current methods involve extracting features from each detected object, which is computationally intensive. We propose VAGNet, a deep neural network that learns to predict accidents from dash-cam video using global features of traffic scenes without requiring explicit object-level features. The network consists of transformer and graph modules, and we use the vision foundation model VideoMAE-V2 for global feature extraction. Experiments on four benchmark datasets (DAD, DoTA, DADA, and Nexar) show that our method anticipates accidents with higher average precision while being computationally more efficient compared to existing methods. It also achieves higher mean time-to-accident scores in most cases, while being competitive in others.
Figures & tables
Figure 1: Architecture of the proposed VAGNet accident anticipation framework.
Dataset
Positive
Negative
FPS
DAD [ 4 ]
620
1130
20
DoTA [ 49 ]
745
745
10
DADA [ 10 ]
938
1361
10
Nexar [ 27 ]
1422
1422
10
Table 1: Dataset Statistics.
Figure 2: Distribution of crash-object categories in the DoTA and DADA datasets. Crash-object category annotations are not available for the Nexar dataset.
Figure 3: Samples showing the variation of predicted accident probability across time/frames for (a) True positive (b) True negative (c) False positive and (d) False negative cases.
Figure 4: Samples of visual explanations for accidents across different weather/lighting conditions. For each condition, the top row shows the original video frames in temporal order, and the bottom row shows the corresponding Grad-CAM [ 33 ] maps highlighting areas in the image that contribute to accident anticipation. Higher intensity indicates regions that have a stronger influence on the model’s anticipation of a potential accident.
Dataset
Method
AP (%)
mTTA (s)
DAD [ 4 ]
DSA (2017) [ 4 ]
48.1
1.34
adaLEA (2018) [ 39 ]
52.3
3.43
UString (2020) [ 1 ]
53.7
3.53
DSTA (2022) [ 14 ]
56.1
3.66
EGSTL (2023) [ 23 ]
56.3
3.23
GSC (2024) [ 48 ]
60.4
3.53
Table 2: Performance comparison of different methods.
Dataset
Global Features
AP (%)
mTTA (s)
DAD
VGG16
54.83
4.15
I3D
59.27
4.13
SlowFast
62.90
3.74
VideoMAE-V2
77.98
4.41
DoTA
VGG16
88.74
3.25
I3D
91.71
3.20
Table 3: Performance of VAGNet under different global feature extraction backbones.
Data
TE
GT
AP (%)
mTTA (s)
DAD [ 4 ]
✗
✗
64.15
2.54
✓
✗
76.58
4.35
✗
✓
71.37
4.12
✓
✓
77.98
4.41
DoTA [ 49 ]
✗
✗
87.50
2.45
✓
✗
95.87
3.31
Table 4: Results of the ablation study. Each model variant removes one or more components (Transformer Encoder or Graph) to analyze their individual impact.
Method
GFLOPs
Object Detection & Feature Extraction
Global Feature Extraction
Model Inference
Total
UString
322.46 (worst case: 19 objects/frame)
15.47 (VGG16)
10.49
348.42
Graph (Graph)
41.75 (I3D)
2.14
366.35
Zhong [ 56 ]
41.75 (I3D)
2.05
366.26
AAT-DA
15.47 (VGG16)
1.05
461.96*
STAGNet
50.58 (SlowFast)
2.20
375.24
Table 5: FLOP counts of different methods. Counts for different stages are reported in GFLOPs/frame. The counts were obtained using the Python library THOP [ 43 ] .
Method
Object Detection & Feature Extraction (s/f)
Global Feature Extraction (s/f)
Inference (s/f)
Overall speed (FPS)
UString
0.032
0.004 (VGG16)
0.024
17
Graph(Graph)
0.01 (I3D)
0.001
23
AAT-DA
0.004 (VGG16)
0.094*
8
STAGNet
0.014 (SlowFast)
0.001
21
VAGNet
0
0.025 (VideoMAE)
0.001
38
Table 6: The computational efficiency of the proposed and baseline methods is evaluated in terms of execution time. Module-level performance is reported in seconds per frame (s/f), whereas overall inference efficiency is reported as frames per second (FPS). For AAT-DA, the reported inference time additionally accounts for the generation of driver attention maps. Note that although we do not report the inference time for the method proposed by Zhong et. al [ 56 ] due to the unavailability of its implementation, we expect it to operate close to 23 FPS on our platform as it uses VGG16 and I3D as feature extractors.
Traffic accident anticipation predicts imminent collisions or rising crash risk early enough to support driver warnings and proactive road-safety decisions.
Anticipating traffic accidents from dashcam videos is a critical challenge in intelligent transportation systems. Existing methods typically map visual context directly to a collision probability without explicitly modeling the future evolution of the driving scene. In this paper we propose FLaRA (Predicting Future Latent Representations for Accident Anticipation), a novel predictive architecture that shifts this paradigm by forecasting future latent representations for accident anticipation. Building upon the Video Joint-Embedding Predictive Architecture (V-JEPA2), our model conditions a predictor network on observed context frames to predict the forthcoming latent features of the scene. A classifier then operates on these predicted future representations rather than only on past observations. To ensure these forecasts remain grounded in realistic future dynamics, we introduce a joint training objective that simultaneously optimizes an auxiliary feature-level reconstruction loss and a cross-entropy classification loss. Extensive evaluations on the Nexar dataset, alongside cross-domain validations on the DAD, DADA-2000, and DoTA benchmarks, demonstrate that our approach achieves state-of-the-art performance while maintaining realistic early warning capabilities.
Lorenzo Caselli, Tomaso Trinci, Tommaso Bianconcini +4
Department of Information Engineering, University of Florence, Italy. · 2Verizon Connect, Florence, Italy.
Traffic accident anticipation -- predicting the likelihood of an imminent collision at every frame of a dashcam video -- is safety-critical yet difficult to scale, because collecting in-domain annotated accident footage for every deployment scenario is prohibitively expensive. We study this task under a zero-shot setting where no target-domain training data is available: the model must learn exclusively from a publicly available binary-labelled driving-accident dataset and generalise to unseen dashcam footage. We propose a framework that bridges the gap between the frame-level temporal risk estimation task and coarsely labelled binary accident datasets by coupling a VideoMAE-v2 backbone with a per-frame prediction head under a sliding-window protocol. Our method achieves 2nd place in the 2026 CVPR@AUTOPILOT Zero-Shot Accident Anticipation competition. Code is available at https://github.com/TimeSouth/zero-shot-taa-solution.
Siyuan Li, Xiaoyang Bi, Mengshi Qi
State Key Laboratory of Networking and Switching Technology · Beijing University of Posts and Telecommunications
Accident anticipation aims to recognize anomalous driving cues before a crash while avoiding false alarms during normal driving. Existing approaches typically formulate this task as binary classification, focusing on whether an accident will occur rather than when it will occur. We propose PRE-ACT, a framework that models accident risk as a continuously evolving signal that increases as the crash approaches. By explicitly enforcing temporal ordering and distance-to-accident awareness, our method progressively raises risk while suppressing premature alarms, leading to significant improvements on MM-AU subsets and Nexar. We further introduce a Separation Score to evaluate the global behavior of predicted risk curves beyond local temporal windows. Code and visualizations are available at https://github.com/giddyyupp/PRE-ACT.
Samet Hicsonmez, Eray Çakar, Nermin Samet +1
Independent Researcher, Paris, France · Department of Computer Engineering, KUIS AI, Koç University, Istanbul, Turkey · Valeo.ai, Paris, France