A Comparative Evaluation of Deep Learning Models for Road Accident Detection Using CCTV Images and Dashcam Videos

Abstract

Road traffic accidents (RTAs) remain a major global safety challenge, requiring reliable automated detection systems to support rapid emergency response, traffic management, and smart city surveillance. This study presents a comparative evaluation of deep learning models for road accident detection using heterogeneous visual traffic sources, namely CCTV images and dashcam videos. A unified evaluation framework is introduced to systematically compare traditional machine learning, standard deep learning, and hybrid deep learning models within a consistent experimental setting. Publicly available datasets comprising several 1000 annotated CCTV images and thousands of dashcam video clips were used, covering diverse traffic scenarios, including urban and highway environments, day and night conditions, weather variations, and multiple camera viewpoints. The evaluated models were grouped into baseline approaches, consisting of Support Vector Machine, 2D Convolutional Neural Network (2D-CNN), R-CNN, and Long Short-Term Memory (LSTM) networks, and a proposed hybrid VGG16-LSTM model. This categorization reflects the complementary strengths of spatial and temporal feature extraction for static image-based and dynamic video-based accident detection. For CCTV image analysis, the 2D-CNN achieved the best performance, with 99% accuracy and a 98% F1-score. For dashcam video analysis, the VGG16-LSTM model outperformed competing approaches, achieving 99.53% accuracy and a high area under the receiver operator characteristic (ROC) curve. The findings demonstrate the effectiveness of convolutional models for spatial accident representation and the importance of temporal modeling for video-based detection. The proposed cross-modality framework provides practical insights into model suitability across different traffic data sources and has potential integration into Intelligent Transportation Systems. Future research should examine cross-dataset validation, event-level detection, and robustness in uncontrolled real-world environments.