Syed Ishmum Ahnaf

Back to all work

Early Accident Anticipation with Video Transformers

Status
In Progress
Period
University thesis — in progress
Confidentiality
Public. Public university thesis repository. Dataset access, third-party terms, and incomplete evaluation status remain visible.

A thesis prototype investigating temporal accident anticipation and fixation-derived supervision with explicit result-integrity safeguards.

I'm exploring early traffic-accident anticipation from dashcam video through a thesis prototype, using temporal sampling, Tiny3D and MViTv2-oriented paths, earliness-aware loss, optional fixation supervision, and sequence evaluation. A result audit found duplicated video rows and conflicting labels, so I'm withholding quantitative claims until the evaluation input is corrected and regenerated.

Type
University
Categories
  • AI & Machine Learning
  • Data & Dashboards
Technologies
  • Python
  • PyTorch
  • torchvision
  • timm
  • NumPy
  • scikit-learn
  • Matplotlib
  • pytest

Context

Researchers and engineers interested in causal video modeling, accident anticipation, and evaluation integrity.

  • The dataset, trained checkpoints, and full ML environment are not included.
  • Committed evaluation summaries contain duplicate keys and conflicting video-level labels.
  • Dataset reuse terms and repository licensing require clarification.

Problem

The research asks whether video models can identify accident risk early and whether fixation-derived supervision provides a useful auxiliary signal, while avoiding future-frame leakage and invalid aggregate conclusions.

My contribution

I built data loading and sampling, Tiny3D and MViTv2-oriented model paths, classification and fixation heads, earliness-aware training, sequence evaluation, threshold and mTTA-related output, deterministic setup, checkpointing, and comparative probability-timeline visualization.

Approach

Resolve video annotations, form causal temporal inputs, encode clips, predict accident likelihood, optionally supervise a fixation distribution, evaluate probability over time, and audit every result artifact before publishing performance conclusions.

Artifacts

Outcomes

  • Implemented the full research prototype and probability-timeline visualization path.
  • Traced repeated evaluation rows to an annotation-row versus video-key granularity mismatch.
  • Withheld unsupported AP, AUC, mTTA, and fixation-benefit claims.

Limitations and current status

  • No quantitative model-performance result is currently publishable.
  • Full tests, training, and plot reproduction require unavailable dependencies, data, and checkpoints.
  • The split parser and time-to-event targets need correction before the next evaluation.