RiskProp: Collision-Anchored Self-Supervised Risk Propagation for Early Accident Anticipation

1. Problem Statement

Accident Anticipation aims to predict impending collisions from dashcam videos in real-time to trigger early alerts.

• Current Limitations: Existing methods rely on binary supervised learning with manually annotated "anomaly onset" frames.
  • Subjectivity: Determining exactly when an accident "starts" (anomaly onset) is subjective and inconsistent across annotators.
  • Inaccuracy: Binary labeling forces all pre-collision frames to be treated as equally risky (label 1) or safe (label 0), ignoring the gradual, intermediate evolution of risk.
  • Consequence: Models are misled to optimize for inaccurate binary targets rather than learning the true, dynamic progression of risk, leading to false alarms or delayed warnings.

2. Methodology: RiskProp

The authors propose RiskProp, a novel collision-anchored self-supervised risk propagation paradigm. Instead of relying on subjective "onset" labels, the method uses only the objectively annotated collision frame as a hard anchor.

Core Concept

The framework models temporal risk evolution through two observation-driven losses:

1. Future-Frame Regularization: Future frames contain more definitive evidence of an accident. Therefore, the model's prediction for the next frame serves as a reliable "soft target" to supervise the current frame.
2. Adaptive Monotonicity: Risk generally follows a non-decreasing trend as a collision approaches. The model is constrained to reflect this physical reality while allowing for short-term fluctuations.

Key Components

A. Labeling Strategy

• Accident Videos: Only the collision frame is labeled as positive ($y=1$). The starting frame (far from the accident) is labeled negative ($y=0$). All intermediate frames are unlabeled.
• Safe Videos: All frames are labeled negative ($y=0$).

B. Loss Functions
The total training objective is: $\mathcal{L} = \mathcal{L}_{\text{bce}} + \lambda_1 \mathcal{L}_{\text{reg}} + \lambda_2 \mathcal{L}_{\text{mono}}$

1. Binary Cross-Entropy (BCE) Loss:

   • Applied only to the collision frame (positive) and the initial frame (negative) in accident videos.
   • Provides the hard anchor for supervision.

2. Future-Frame Regularization Loss ($\mathcal{L}_{\text{reg}}$):

   • Mechanism: Uses the model's prediction for the next frame ($z_{t+1}$), with gradients detached, as a soft target for the current frame ($z_t$).
   • Formula: $\mathcal{L}_{\text{reg}} = \sum_{t=1}^{T-1} \| \text{detach}(z_{t+1}) - z_t \|^2$
   • Effect: Enables backward propagation of risk signals. The high risk score at the collision frame propagates backward through the sequence, teaching the model to recognize early cues (e.g., lane drifts, delayed reactions) without explicit onset labels.

3. Adaptive Monotonic Constraint Loss ($\mathcal{L}_{\text{mono}}$):

   • Mechanism: Penalizes cases where a later frame is predicted to be less risky than an earlier frame ($a_j < a_i$ for $j > i$).
   • Adaptive Margin: The constraint strength is not fixed. It adapts based on:
     • Temporal Distance ($\Delta t$): Larger gaps allow for more expected risk increase.
     • Prediction Confidence: Stricter constraints are applied when the model is confident (scores near 0 or 1); looser constraints are applied for ambiguous intermediate states.
   • Formula: $\mathcal{L}_{\text{mono}} = \frac{1}{|\mathcal{D}|} \sum_{(i,j) \in \mathcal{D}} \max(0, a_i - a_j + \delta(\Delta t, \bar{c}_{i:j}))$

C. Architecture

• Uses a 3D CNN (specifically SlowOnly) as the snippet encoder to extract spatial-temporal features.
• Processes video snippets (e.g., 5 frames) and outputs a risk score for the current frame.

3. Key Contributions

1. Paradigm Shift: Moves away from subjective "anomaly onset" annotations to a collision-anchored self-supervised approach.
2. Risk Propagation: Introduces a mechanism where risk signals propagate backward from the definitive collision point to earlier frames using future-frame regularization.
3. Adaptive Constraints: Proposes an adaptive monotonic loss that enforces a physically plausible, non-decreasing risk trend while accommodating short-term noise and uncertainty.
4. Annotation Efficiency: Demonstrates that high-performance accident anticipation can be achieved using only collision labels, eliminating the need for costly and subjective onset annotations.

4. Experimental Results

The method was evaluated on two real-world datasets: CAP (11,727 videos) and Nexar (1,500 clips).

• State-of-the-Art Performance:
  • CAP Dataset: Achieved mAUC0.1 of 0.483 and mTTA0.1 of 1.207s (earliest warning), outperforming all baselines (e.g., AdaLEA, DSTA, CRASH).
  • Nexar Dataset: Achieved mAUC0.1 of 0.472, mAP of 0.870, and mTTA0.1 of 0.958s, surpassing the second-best method (CRASH) by significant margins.
• Ablation Studies:
  • Removing the self-supervised constraints (using only collision labels) resulted in poor performance.
  • Adding Future-Frame Regularization (FFR) provided the largest performance boost, effectively propagating risk signals.
  • Adding Adaptive Monotonic Constraint (AMC) further refined the curves, improving temporal consistency.
  • The full model with Only Collision supervision matched or exceeded models trained with dense "anomaly onset" labels.
• Qualitative Analysis:
  • RiskProp produces smoother, more discriminative risk curves.
  • Unlike baselines that trigger false alarms during ambiguous safe periods (e.g., a car appearing on the side), RiskProp maintains low risk scores until clear danger cues emerge, then rises sharply.

5. Significance

• Practicality: By removing the need for subjective "anomaly onset" annotations, RiskProp significantly reduces the cost and difficulty of data preparation for accident anticipation systems.
• Reliability: The self-supervised approach yields risk curves that are more physically plausible and temporally consistent, reducing false positives and enabling earlier, more reliable warnings.
• Generalizability: The framework shifts the focus from static classification to dynamic risk evolution modeling, offering a robust solution for real-world autonomous driving and driver assistance systems.