A Strategy-Driven Training Pipeline for Stable Traffic Accident Anticipation via Cross-Dataset Motion Transfer and Progressive Supervision

Abstract

Vision-based traffic accident anticipation is critical for active vehicle safety systems, yet existing architectures frequently conflate performance gains with heavy parameter scaling optimized from scratch on compact domains. Consequently, during real-time inference, these frameworks suffer from severe prediction volatility and early triggering biases that induce dangerous control instability. To address these limitations, this paper shifts the research focus away from network modifications toward a highly controlled, strategy-driven training pipeline executed under a completely invariant spatial–temporal neural backbone. Our proposed paradigm establishes a robust framework through three decoupled milestones. First, an out-of-domain initialization strategy transferred generalized driving kinetics from a large-scale sequence domain (Mapillary) to serve as a stable temporal anchor. Second, a target-domain generative enrichment step injected synthetic nighttime scenes to decouple hazard features from low-light ambient noise. Third, progressive temporal supervision paradigm scaling targeted labels monotonically to align with continuous kinetic risk accumulation. Overall evaluations on the Car Crash Dataset (CCD) benchmark demonstrate that the fully integrated configuration (C4) pipeline achieves 69.89% in frame-level Mean Average Precision (mAP), which is an improvement of +22.81 percentage points over the baseline configuration. Continuous temporal measurements prove that our framework can adapt to tracking volatility, compressing Temporal Confidence Variance to 0.00328, and dropping the Prediction Instability Count to 0.66. While hyper-sensitive baselines report early raw latency averages driven by premature trigger noise, our model purposefully filters this early-frame variability to deliver a secure warning profile, achieving an absolute zero false alarm rate (FAR = 0.00%) across evaluated non-hazardous driving sequences, establishing the sequence-level trustworthiness required for practical autonomous deployment.