Research map / Detection and segmentation / 3D detection and driving
DeepAccident: A Motion and Accident Prediction Benchmark for V2X Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DeepAccident, a CARLA-simulated V2X dataset of intersection collision scenarios with an end-to-end motion and accident prediction task, and a V2XFormer baseline that outperforms a single-vehicle BEVerse model.
Problem
Autonomous driving lacks datasets that support direct, explainable safety evaluation via accident prediction, and existing V2X datasets lack motion/accident labels and safety-critical collision scenarios.
Approach
A large simulator-generated V2X dataset of 12 NHTSA-pre-crash-inspired intersection accident types recorded from four vehicles and one infrastructure, plus an end-to-end camera-based motion and accident prediction task (APA metric) and a V2XFormer model that fuses BEV features across agents.
Key results
- V2XFormer with four vehicles plus infrastructure reaches 56.2 mIOU, 44.0 VPQ, 69.5 APA and 40.8 mAP on validation, versus 43.8 mIOU, 31.6 VPQ, 61.9 APA and 26.5 mAP for the single-vehicle model.
- On invisible accident-vehicle samples, the 5-agent model beats the single-vehicle model by 15.9 mIOU and 12.2 APA.
- Fine-tuning on DeepAccident then nuScenes improves nuScenes VPQ from 33.4 to 34.2 and mAP from 32.1 to 34.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Average Fusion with CoBEVT fusion (ablation) | V2XFormer · DeepAccident validation | APA: 67.1 -> 69.5 | Helped |
| Add V2X communication using four vehicles and infrastructure | Single-vehicle BEVerse-tiny · DeepAccident validation | mIOU: 43.8 -> 56.2 | Helped |
| Add V2X communication using four vehicles and infrastructure | Single-vehicle BEVerse-tiny · DeepAccident validation | APA: 61.9 -> 69.5 | Helped |
| Add V2X communication using four vehicles and infrastructure | Single-vehicle BEVerse-tiny · DeepAccident validation | mAP: 26.5 -> 40.8 | Helped |
| Add training on DeepAccident before nuScenes fine-tuning | BEVerse-tiny · nuScenes validation | VPQ: 33.4 -> 34.2 (+0.8) | Helped |
| Add training on DeepAccident before nuScenes fine-tuning | BEVerse-tiny · nuScenes validation | mAP: 32.1 -> 34.0 (+1.9) | Helped |
| Extend prediction horizon from 2s to 3s (ablation) | Single-vehicle model · DeepAccident validation, all data | APA: 61.9 -> 50.5 | Hurt |
| Extend prediction horizon from 2s to 4s (ablation) | Single-vehicle model · DeepAccident validation, all data | APA: 61.9 -> 35.4 | Hurt |
| Extend prediction horizon to 4s (ablation) | Single-vehicle model · DeepAccident validation, 4s before collision | APA: 10.2 | Helped |
Surprising findings
- Predicting longer horizons sharply reduces accident prediction accuracy: 4s-horizon APA is 35.4 versus 61.9 for the 2s model on all validation data.
Limitations
- Accident scenarios are limited to 12 intersection types generated in simulation, and sim-to-real gap is only partially checked via nuScenes fine-tuning.
- Longer prediction horizons degrade accident prediction accuracy, and more V2X agents increase communication bandwidth and latency.
- Accident prediction relies on post-processing of motion outputs with a 1.0 m danger threshold rather than a dedicated predictor.
Experimental setup
Datasets: DeepAccident nuScenes
Benchmarks: DeepAccident nuScenes
Metrics: mIOU VPQ APA id error position error mAP
Baselines: BEVerse-tiny (single-vehicle) Average Fusion DiscoNet V2X-ViT CoBEVT
Builds on: BEVerse FIERY CoBEVT CARLA
Closest papers by meaning
- DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles (AAAI 2025 · similarity 0.85)
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · similarity 0.85)
- SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.84)
- Impromptu VLA: Open Weights and Open Data for Driving Vision-Language-Action Models (NeurIPS 2025 · similarity 0.84)
- Learning Cooperative Trajectory Representations for Motion Forecasting (NeurIPS 2024 · similarity 0.83)
- End-to-End Autonomous Driving Through V2X Cooperation (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.