A Hybrid Lightweight Framework for Real-Time CCTV-based Accident Detection with Automated Emergency Alert Generation

Abstract

The high rate of growth in road traffic has been a major cause of drastic increase in the number of accidents which have resulted in high demand of effective real-time monitoring and response systems. This paper introduces a hybrid system of accident detection and classification of severity based on CCTV video streams implemented on deep learning. The suggested system is a combination of the YOLOv8s object detector model and a Convolutional Neural Network (CNN) using MobileNetV3 to guarantee the efficient performance of the system. YOLOv8s is used to identify regions of accidents in the continuous video frames and the CNN model is used to categorize the identified accidents into moderate and severe. The system is trained on a publicly available dataset of more than 15,000 annotated images, which were recorded in a variety of environmental conditions, which guarantees a robust system and generalization. Experimental findings indicate that it has a high detection accuracy with good precision, recall and mAP values and stable classification accuracy. Also, an automated alert system is included to inform emergency services in case of severe accidents identified so that they can respond quicker and have a lesser effect. The suggested framework is minimal and scalable and appropriate to be implemented in intelligent transportation systems and smart city applications.