RADAR: Road Accident Detection and Analysis using Recognition

Abstract

Road traffic accidents are a major cause of fatalities and economic loss worldwide, necessitating efficient and automated monitoring solutions. Manual surveillance of CCTV footage is labor-intensive, error-prone and unsuitable for continuous large-scale monitoring. This work presents RADAR, a deep learning framework designed to automatically identify road accidents from video-based surveillance data. The framework combines both spatial and temporal learning mechanisms to capture visual features and motion patterns associated with accident events. Spatial features are extracted from individual video frames using a pre-trained VGG16 convolutional neural network, enabling robust representation of vehicles, road structures and collision regions. To model temporal dynamics across consecutive frames, a stacked Gated Recurrent Unit (GRU) architecture is employed for sequence-level classification. Considering the rarity of accident events and the associated class imbalance problem, Time-Series Generative Adversarial Networks (TimeGAN) are utilized to generate synthetic accident sequences, enhancing dataset diversity and improving generalization performance. Experimental evaluation demonstrates that the proposed CNN–GRU hybrid architecture effectively distinguishes accident and non-accident sequences, achieving high classification accuracy and improved robustness against limited training samples. The framework presents a scalable foundation for intelligent traffic surveillance systems and contributes toward AI-driven smart transportation infrastructure.