From Chaos to Detection: Accident Benchmarking in Surveillance Videos with a Curated Dataset and 3D CNNs

Received: 18 December 2025 | Revised: 31 January 2026, 4 February 2026, and 7 February 2026 | Accepted: 8 February 2026 | Online: 6 June 2026

Corresponding author: Ranjit Singh Sarban Singh

Abstract

Automatic accident detection from surveillance videos is an important task for intelligent transportation and public safety, yet it remains underdeveloped compared to violence and anomaly detection. While existing methods often report high accuracy, these results are frequently based on contaminated datasets containing duplicates, overlapping scenes, or annotation artifacts, which inflate performance and limit real-world applicability. To address this gap, this paper introduces a curated benchmark dataset of 513 videos from multiple sources, which are segmented into 4 s clips, resulting in 1,122 clips (561 per class). Using this dataset, we evaluate a spectrum of approaches ranging from handcrafted Violent Flows (ViF) + Support Vector Machine (SVM) pipelines to Convolutional Neural Network–Recurrent Neural Network (CNN–RNN) hybrids and modern 3D convolutional networks. The conducted experiments confirm that traditional methods collapse under stricter evaluation, whereas lightweight architectures such as X3D achieve the best balance between accuracy and efficiency. The reconfigured X3D-S variant achieved 84.48% accuracy, establishing a strong baseline for accident detection under realistic conditions, while offering both a cleaner benchmark and a practical design for future deployment. The Surveillance Curated Accident Dataset (SCAD) and full implementation code are publicly available and can be cited through this paper.