Nexar Collision Prediction Challenge

Event: CVPR 2025 Workshop on Autonomous Driving · Duration: 19 min · ▶ Watch on YouTube

Abstract

The Nexar Collision Prediction Challenge aims to advance research in anticipating traffic accidents using road-facing dashcam footage. Nexar, with its vast collection of over 60 million videos from 350,000 dashcams, released a high-quality open dataset of 1500 videos featuring near-collisions and normal driving scenarios, meticulously annotated with temporal events like ‘Time of Alert’ and ‘Time of Collision’. The challenge introduced a novel mean Average Precision (mAP) metric to unify the goals of early prediction, minimizing false positives, and maximizing recall across different anticipation levels. Analysis of top solutions revealed the effectiveness of video transformers and optical flow-based methods, while highlighting the increased difficulty of predicting accidents further in advance and in challenging conditions like fog.

Speakers

  • Daniel Moura — Nexar

Talks (1)

  • 00:04 — Daniel Moura: Nexar Collision Prediction Challenge
    • This talk introduces the Nexar Collision Prediction Challenge, detailing the Nexar Open Dataset of dashcam footage with precise temporal annotations for collision-related events, the proposed mAP metric for evaluating early and accurate predictions, and insights from top-performing solutions.

Key Takeaways

  • Introduced ‘alert time’ as a crucial temporal annotation for collision-related events, capturing the moment an action should be taken to avoid an accident.
  • Proposed ‘mean Average Precision (mAP)’ as a unified metric for collision anticipation, effectively combining early prediction, false positive minimization, and recall maximization.
  • Released a comprehensive open dataset of 1500 high-definition dashcam videos with detailed annotations and evaluation scripts, making it ready for community research.
  • Gained insights into key solutions from the community, with video transformers (MViT, VideoMAE) and optical flow-based methods showing promising results, and identified challenging scenarios like fog and longer anticipation times.

Methods / Models / Datasets Mentioned

  • MViTv2-S
  • VideoMAEv2
  • ResNet18
  • Yolo v8
  • GRU

Topics

Collision anticipation · Dashcam data · Open dataset · Temporal annotation · Mean Average Precision (mAP) · Video transformers · Optical flow · Traffic accidents · Autonomous driving

Notes

Open for commentary — connections to other work, critiques, follow-up reading.