Research map / Video / Video segmentation and tracking
Accident Anticipation via Temporal Occurrence Prediction
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.
Shifts accident anticipation from per-frame risk scoring to predicting accident scores at multiple future horizons (0.1–2.0s) with a snippet-level 3D encoder and Transformer temporal decoder, and proposes FAR-constrained recall/TTA evaluation.
Problem
Existing accident anticipation methods predict frame-level risk scores trained with ambiguous binary supervision (all accident frames labeled positive), causing unreliable learning and false alarms; evaluation metrics also ignore false alarm rates, making recall and TTA misleading.
Approach
Predict a sequence of accident-occurrence scores at multiple future time steps using temporal queries in a Transformer decoder, supervised by the precise accident timestamp (single positive per snippet), on top of a snippet-level 3D CNN encoder; evaluate recall and TTA only under a FAR threshold lambda, and compute TTA only after the anomaly appears.
Key results
- On CAP with FAR<=10% (lambda=0.1), the method reaches mAUC 0.4290 and mTTA 0.8644 s, versus best prior GSC at mAUC 0.3046 and mTTA 0.8165 s.
- On DADA with lambda=0.1, mAUC 0.3315 and mTTA 0.8848 s, versus best prior GSC at mAUC 0.2423.
- Ablation on CAP: adding both TOP and snippet encoder yields AUC^0.1 0.8381 vs 0.4357 for the frame-level baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace frame-level risk scoring with temporal occurrence prediction and snippet-level encoding | Accident anticipation model · CAP | mAUC: 0.3046 -> 0.4290 | Helped |
| replace frame-level risk scoring with temporal occurrence prediction and snippet-level encoding | Accident anticipation model · DADA | mAUC: 0.2423 -> 0.3315 | Helped |
| add the snippet encoder without temporal occurrence prediction (ablation) | Baseline accident anticipation model · CAP | AUC (0.5s): 0.3938 -> 0.5550 | Helped |
| add the snippet encoder without temporal occurrence prediction (ablation) | Baseline accident anticipation model · CAP | mTTA (s): 0.5389s -> 0.7330s | Helped |
| add temporal occurrence prediction without the snippet encoder (ablation) | Baseline accident anticipation model · CAP | AUC (0.0s): 0.4357 to 0.5700 | Helped |
| add temporal occurrence prediction without the snippet encoder (ablation) | Baseline accident anticipation model · CAP | AUC (0.5s): 0.3938 -> 0.3432 | Hurt |
| add the snippet encoder to temporal occurrence prediction (ablation) | Experiment III (TOP only) · CAP | mAUC: 0.2479 -> 0.4290 | Helped |
| add the snippet encoder to temporal occurrence prediction (ablation) | Experiment III (TOP only) · CAP | mTTA (s): 0.4595 -> 0.8644 | Helped |
Surprising findings
- Previously reported TTAs exceeding 3 seconds are argued to be artifacts of counting false alarms before the anomaly as valid early predictions.
- TOP alone without the snippet encoder gives limited gains at longer horizons, with mAUC 0.2479 below the baseline's 0.2829 on CAP.
Limitations
- Performance at longer horizons (>1.0s) remains limited.
- Spurious alerts still occur in complex scenes even under constrained FAR.
- Anomaly-appear annotations are sometimes inconsistent, adding label noise.
Experimental setup
Datasets: MM-AU CAP DADA CCD A3D DoTA DADA-2000
Benchmarks: CAP DADA
Metrics: AUC^lambda mAUC^lambda mTTA^lambda recall at 0.5s/1.0s/1.5s pre-accident intervals
Baselines: CAP DRIVE DSTA GSC ResNet+LSTM frame-level risk baseline
Builds on: SlowOnly Transformer decoder
Closest papers by meaning
- Progressive Risk Estimation for Accident Anticipation (NeurIPS 2026 · similarity 0.92)
- RiskProp: Collision-Anchored Self-Supervised Risk Propagation For Early Accident Anticipation (CVPR 2026 · similarity 0.90)
- Anticipating Traffic Accidents With Adaptive Loss and Large-Scale Incident DB (CVPR 2018 · similarity 0.90)
- Harnessing Event Sensory Data for Error Pattern Prediction in Vehicles: A Language Model Approach (AAAI 2025 · similarity 0.86)
- TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents (NeurIPS 2025 · similarity 0.85)
- DeepAccident: A Motion and Accident Prediction Benchmark for V2X Autonomous Driving (AAAI 2024 · similarity 0.82)
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.