BADAS: Context Aware Collision Prediction Using Real-World Dashcam Data

TL;DR

BADAS addresses the mismatch between collision-prediction data and real-world ego-centric safety by reframing the problem to ego-centric threats and re-annotating major benchmarks for ego involvement and consensus alert timing. It leverages the V-JEPA2 video foundation model, fine-tuned on real ego-centric dashcam data, and introduces two model variants, BADAS-Open and BADAS1.0, to study data-scaling effects. The results show state-of-the-art AP/AUC and more realistic mean time-to-accident estimates, demonstrating the practical benefits of ego-centric training, standardized timing, and synthetic negatives. The work emphasizes the importance of real-world data, temporal evaluation standards, and long-tail analysis for deploying reliable collision-prediction systems in ADAS and autonomous driving contexts.

Abstract

Existing collision prediction methods often fail to distinguish between ego-vehicle threats and random accidents not involving the ego vehicle, leading to excessive false alerts in real-world deployment. We present BADAS, a family of collision prediction models trained on Nexar's real-world dashcam collision dataset -- the first benchmark designed explicitly for ego-centric evaluation. We re-annotate major benchmarks to identify ego involvement, add consensus alert-time labels, and synthesize negatives where needed, enabling fair AP/AUC and temporal evaluation. BADAS uses a V-JEPA2 backbone trained end-to-end and comes in two variants: BADAS-Open (trained on our 1.5k public videos) and BADAS1.0 (trained on 40k proprietary videos). Across DAD, DADA-2000, DoTA, and Nexar, BADAS achieves state-of-the-art AP/AUC and outperforms a forward-collision ADAS baseline while producing more realistic time-to-accident estimates. We release our BADAS-Open model weights and code, along with re-annotations of all evaluation datasets to promote ego-centric collision prediction research.

Figures (9)

• Figure 1: BADAS achieves state-of-the-art performance on collision prediction benchmarks. Our approach substantially outperforms existing academic methods and commercial vision-based forward collision warning systems by leveraging ego-centric real-world data and modern video foundation models when evaluated on ego-vehicle involved collisions.
• Figure 2: The critical distinction between ego-centric and general collision prediction. Top row: Ego-vehicle involved events requiring immediate driver intervention. Bottom row: Non-ego accidents that are visually prominent but irrelevant to ego-vehicle safety. BADAS focuses exclusively on ego-relevant scenarios to minimize false alarms while maintaining high sensitivity to actual threats. Examples from the DAD dataset chan2017anticipating.
• Figure 3: Cumulative distribution of human reaction times across 726 ego-involved collisions. Median: 1.70s, 90% range: 0.70-3.47s before impact.
• Figure 4: Collision predictions scores over time for 6 samples: UString (blue), DSTA (green) and BADAS-Open (red). Vertical dashed lines indicate collision time, horizontal line shows detection threshold.
• Figure 5: Frame-by-frame comparison on Nexar test footage. Top: urban intersection collision. Bottom: residential street scenario. Alert indicators show prediction confidence (green: safe, orange: caution, red: imminent collision). BADAS variants demonstrate accurate ego-centric threat assessment.