Organizations: School of Computer Science, Wuhan University · Zhongguancun Academy, Beijing, China · Didi Chuxing · University of Science and Technology of China
Accident anticipation aims to predict impending collisions from dashcam videos and trigger early alerts. Existing methods rely on binary supervision with manually annotated "anomaly onset" frames, which are subjective and inconsistent, leading to inaccurate risk estimation. In contrast, we propose RiskProp, a novel collision-anchored self-supervised risk propagation paradigm for early accident anticipation, which removes the need for anomaly onset annotations and leverages only the reliably annotated collision frame. RiskProp models temporal risk evolution through two observation-driven losses: first, since future frames contain more definitive evidence of an impending accident, we introduce a future-frame regularization loss that uses the model's next-frame prediction as a soft target to supervise the current frame, enabling backward propagation of risk signals; second, inspired by the empirical trend of rising risk before accidents, we design an adaptive monotonic constraint to encourage a non-decreasing progression over time. Experiments on CAP and Nexar demonstrate that RiskProp achieves state-of-the-art performance and produces smoother, more discriminative risk curves, improving both early anticipation and interpretability.
Figures & tables
Figure 1: Form of supervision of previous works and ours. (a) Most previous methods treat all frames in accident-free videos as negative samples with label 0, and treat frames in accident videos—from the annotated anomaly onset frame to the collision frame—as positive samples with labeled 1. (b) Our method treats the model’s prediction for the next frame as a supervision signal for the current frame, which enables the risk values to propagate gradually backward from the collision frame.
Figure 2: Overview of our RiskProp framework. The encoder-only model takes a snippet of consecutive frames and predicts the current frame’s risk score. To train such a model without “anomaly onset” labels and with only the objective collision frame label, two losses are proposed: The Future-Frame Regularization Loss uses the next-frame’s detached prediction as a self-supervised target for high-risk signals backward propagation, while the Adaptive Monotonic Constraint Loss imposes a monotonicity constraint to ensure a non-decreasing overall risk trend. And we adopt the Binary Cross-Entropy loss to provide explicit supervision only at the collision frame. This enables stable, physically plausible risk curves without manual onset annotations.
Figure 3: Illustration of sampling strategy for the adaptive monotonic constraint loss. For a randomly selected d∈[dmin,dmax] , a starting frame t1 is sampled from [0,T(1−d)] , and t2=t1+dT . This encourages learning across various temporal distances.
Method
CAP [ 8 ]
Nexar [ 26 ]
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
AdaLEA [ 33 ]
0.379
0.807
0.857
1.115
0.378
0.828
0.832
0.858
XAI [ 18 ]
0.352
0.821
0.864
1.082
0.346
0.824
0.802
0.839
DSTA [ 17 ]
0.361
0.895
0.882
0.894
0.242
0.783
0.764
0.493
GSC [ 36 ]
0.378
0.881
0.890
0.911
0.322
0.802
0.811
0.815
CAP [ 8 ]
0.332
0.811
0.852
0.933
0.315
0.817
0.793
0.801
Table 1: Quantitative results comparison of different methods on the CAP [ 8 ] dataset and the Nexar [ 26 ] dataset.
Exp.
FFR
AMC
Anomaly Onset Label
Fixed Interval Label
Only Collision Label
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
I
0.441
0.829
0.871
1.162
0.412
0.816
0.872
1.173
0.358
0.783
0.832
1.009
II
✓
0.436
0.828
0.866
1.137
0.436
0.828
0.872
1.173
0.383
0.790
0.832
1.055
III
✓
0.444
0.830
0.872
1.174
0.455
0.847
0.886
1.212
0.474
0.850
0.850
1.202
IV
✓
✓
0.484
0.856
0.887
1.198
0.480
0.851
0.890
1.203
0.483
0.853
0.890
1.207
Table 2: Ablation study about different annotation strategies on the CAP [ 8 ] dataset, where FFR denotes Future-Frame Regularization, and AMC denotes Adaptive Monotonic Constraint.
Exp.
FFR
AMC
Anomaly Onset Label
Fixed Interval Label
Only Collision Label
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
mAUC 0.1
mAUC
mAP
mTTA 0.1 (s)
I
0.388
0.826
0.822
0.773
0.420
0.837
0.827
0.741
0.298
0.789
0.781
0.610
II
✓
0.408
0.844
0.842
0.902
0.445
0.815
0.830
0.815
0.302
0.797
0.785
0.654
III
✓
0.434
0.846
0.850
0.897
0.434
0.865
0.856
0.920
0.453
0.847
0.854
0.836
IV
✓
✓
0.479
0.872
0.876
0.951
0.454
0.875
0.874
0.931
0.472
0.869
0.870
0.958
Table 3: Ablation study about different annotation strategies on the Nexar [ 26 ] dataset, where FFR denotes Future-Frame Regularization, and AMC denotes Adaptive Monotonic Constraint.
Figure 4: Frame- and Dataset-Level Risk Prediction. Left: On a representative accident video, supervised baselines produce early false peaks before clear risk emerges. Our RiskProp model prediction remains low during safe periods and rises sharply only when discriminative cues appear, yielding a temporally coherent risk curve. Right: Dataset-wide average risk curves on CAP and Nexar, aligned to accident timestamps. RiskProp suppresses early false positives and delivers a steeper, better-calibrated rise near the event, enabling earlier, more reliable warnings.