Predict and Resist: Long-Term Accident Anticipation under Sensor Noise

TL;DR

Addresses the challenge of accident anticipation under noisy sensor inputs while requiring timely warnings. Proposes a unified framework that fuses diffusion-based denoising with a time-aware actor-critic for long-horizon risk forecasting, enhanced by state-history processing and dual-level feature refinement. Key contributions include framing accident anticipation as a long-horizon credit assignment problem, image- and object-level diffusion modules for noise-robust features, a time-weighted anticipation loss and actor-critic objective, and state-of-the-art AP and mean Time-to-Accident ($\text{mTTA}$) on DAD, CCD, and A3D under Gaussian and impulse noise. Qualitative analyses show earlier, more stable, human-aligned predictions in routine and complex traffic, supporting strong potential for real-world safety deployment. Overall, the work advances proactive autonomous driving by enabling robust, early warnings in degraded sensing environments.

Abstract

Accident anticipation is essential for proactive and safe autonomous driving, where even a brief advance warning can enable critical evasive actions. However, two key challenges hinder real-world deployment: (1) noisy or degraded sensory inputs from weather, motion blur, or hardware limitations, and (2) the need to issue timely yet reliable predictions that balance early alerts with false-alarm suppression. We propose a unified framework that integrates diffusion-based denoising with a time-aware actor-critic model to address these challenges. The diffusion module reconstructs noise-resilient image and object features through iterative refinement, preserving critical motion and interaction cues under sensor degradation. In parallel, the actor-critic architecture leverages long-horizon temporal reasoning and time-weighted rewards to determine the optimal moment to raise an alert, aligning early detection with reliability. Experiments on three benchmark datasets (DAD, CCD, A3D) demonstrate state-of-the-art accuracy and significant gains in mean time-to-accident, while maintaining robust performance under Gaussian and impulse noise. Qualitative analyses further show that our model produces earlier, more stable, and human-aligned predictions in both routine and highly complex traffic scenarios, highlighting its potential for real-world, safety-critical deployment.

Figures (3)

• Figure 1: Overview illustration of our framework. The figure highlights the integration of the diffusion module and reinforcement learning for processing noisy input scenes, leading to earlier and more reliable accident anticipation.
• Figure 2: Overview of the proposed framework. Input frames are encoded into image and object features, refined by object-aware and diffusion modules. Fused features are processed by a GRU with time-weighted layers to predict $p_t$, while an actor-critic module optimizes long-horizon early warnings.
• Figure 3: Visualization of long-horizon vs. frame-level anticipation on DAD (threshold 0.5). Scenarios: (a) ambiguous multi-agent rain, (b) predictable collision, (c) sudden complex crash. The long-horizon model offers earlier, more reliable predictions with fewer false positives.