A Hierarchical Event-then-Severity Cascade for Crash-severity Grading from Crowdsourced Dashcam Video

Video-based road-crash analysis has focused on binary accident detection, leaving crash severity largely unaddressed, as few dashcam datasets annotate it.We introduce GraDE (Graded Dashcam Events), a source-pure corpus of 1,321 dashcam clips from 25 compilations, labelled non-crash, near-miss, and crash, with a minor/major severity grade on 433 human-audited crash clips.As a reference system for this task we build a hierarchical eventthen-severity cascade from established components: a VideoMAE event head routes predicted crashes to a severity stage fusing a second VideoMAE head with a ResNet-50 + BiLSTM head by late fusion.Under source-pure four-fold cross-validation the event head reaches macro-F1 0.8951 (0.8847 averaged over three seeds), above re-implemented CNN (0.8655) and CNN + BiLSTM (0.8686) baselines, and fused severity reaches 0.6730.Retraining on clips with frame corners masked, borders cropped, resolution reduced and colour standardised lowers event macro-F1 to about 0.72, showing that the event head is sensitive to frame-edge and appearance information.A modern MViTv2-S baseline scores higher on events (0.9161 against 0.8924, both three-seed ensembles) and, as the event stage, lifts endto-end macro-F1 to 0.7552 from 0.7310 (three-seed means), so the cascade is presented as backbone-agnostic.Late fusion, with a VideoMAE fallback for clips routed in error, lifts end-to-end four-class macro-F1 from 0.7239 to 0.7428; the gain, which we treat as exploratory, is positive in three of four folds and for all three Stage A seeds and with either event backbone, but is not statistically established under source-clustered bootstrap.Zero-shot transfer to the Car Crash Dataset provides external evidence that VideoMAE's relative advantage over the CNN baselines is not confined to GraDE.This transfer tests event recognition only; the severity result, which holds across held-out crash sources, is an initial benchmark on GraDE.

Authors

Institutions

Publication Details

Journal
International journal of intelligent engineering and systems
Published
2026-11-17
DOI
https://doi.org/10.22266/ijies2026.1231.34
Primary Topic
Human Pose and Action Recognition
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Hierarchical Event-then-Severity Cascade for Crash-severity Grading from Crowdsourced Dashcam Video

Nikhil N. Gondaliya, Mittal Darji
1 links
International journal of intelligent engineering and systems
Human Pose and Action Recognition
article

A Hierarchical Event-then-Severity Cascade for Crash-severity Grading from Crowdsourced Dashcam Video

Nikhil N. Gondaliya, Mittal Darji
article en

Abstract

Video-based road-crash analysis has focused on binary accident detection, leaving crash severity largely unaddressed, as few dashcam datasets annotate it.We introduce GraDE (Graded Dashcam Events), a source-pure corpus of 1,321 dashcam clips from 25 compilations, labelled non-crash, near-miss, and crash, with a minor/major severity grade on 433 human-audited crash clips.As a reference system for this task we build a hierarchical eventthen-severity cascade from established components: a VideoMAE event head routes predicted crashes to a severity stage fusing a second VideoMAE head with a ResNet-50 + BiLSTM head by late fusion.Under source-pure four-fold cross-validation the event head reaches macro-F1 0.8951 (0.8847 averaged over three seeds), above re-implemented CNN (0.8655) and CNN + BiLSTM (0.8686) baselines, and fused severity reaches 0.6730.Retraining on clips with frame corners masked, borders cropped, resolution reduced and colour standardised lowers event macro-F1 to about 0.72, showing that the event head is sensitive to frame-edge and appearance information.A modern MViTv2-S baseline scores higher on events (0.9161 against 0.8924, both three-seed ensembles) and, as the event stage, lifts endto-end macro-F1 to 0.7552 from 0.7310 (three-seed means), so the cascade is presented as backbone-agnostic.Late fusion, with a VideoMAE fallback for clips routed in error, lifts end-to-end four-class macro-F1 from 0.7239 to 0.7428; the gain, which we treat as exploratory, is positive in three of four folds and for all three Stage A seeds and with either event backbone, but is not statistically established under source-clustered bootstrap.Zero-shot transfer to the Car Crash Dataset provides external evidence that VideoMAE's relative advantage over the CNN baselines is not confined to GraDE.This transfer tests event recognition only; the severity result, which holds across held-out crash sources, is an initial benchmark on GraDE.

International journal of intelligent engineering and systemsVol. 19(12)
The Charutar Vidya Mandal (CVM) University (IN)
Human Pose and Action Recognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon