Szczegóły publikacji
Opis bibliograficzny
AcciDetect: a graph-recurrent framework for detecting road accidents from video sequences / Syed Muhammad ABRAR AKBER, Sadia Nishat Kazmi, Ali Muqtadir, Xi Vincent Wang // Journal of Computational Science ; ISSN 1877-7503 . — 2026 — vol. 100 art. no. 102959, s. 1–10. — Bibliogr. s. 10, Abstr. — Publikacja dostępna online od: 2026-07-13
Autorzy (4)
- AGHAbrar Akber Syed Muhammad
- Kazmi Sadia Nishat
- Muqtadir Ali
- Wang Xi Vincent
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 170007 |
|---|---|
| Data dodania do BaDAP | 2026-09-15 |
| Tekst źródłowy | URL |
| DOI | 10.1016/j.jocs.2026.102959 |
| Rok publikacji | 2026 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Creative Commons | |
| Czasopismo/seria | Journal of Computational Science |
Abstract
Accident detection is an important element of modern intelligent transportation systems, providing timely warning to reduce accident risks and improve road safety. However, the majority of existing accident detection models often face challenges in balancing precision with early detection. Attention-based approaches usually achieve moderate precision, while recurrent models are limited in capturing complex dependencies between objects. These challenges lead to unstable and inconsistent performance in different traffic scenarios. To address these challenges, we propose AcciDetect, a graph-recurrent framework for video-based road accident detection that combines Graph Convolutional Networks (GCN) with Bidirectional Gated Recurrent Units (BiGRU). The model learns from object-level and motion-aware representations extracted from sampled frames of dashcam video streams, enabling it to capture both spatial interactions among road users and their temporal evolution for accident detection. We evaluate AcciDetect on two well-known datasets, the Dashcam Accident Dataset (DAD) and the Car Crash Dataset (CCD). We evaluate and validate the effectiveness of AcciDetect through a comparative analysis with existing state-of-the-art models. The evaluation shows that AcciDetect outperforms the existing models. We also substantiated our evaluation with an ablation study that demonstrates the synergistic benefits of GCN and BiGRU. The ablation study confirms that the integration of GCN with BiGRU leads to significant performance improvements compared to purely spatial or purely temporal versions. These results confirm the proposed AcciDetect as a robust and generalizable solution for the detection of traffic accidents under different traffic conditions.
