AI-Based Road Accident Detection and Prediction Using Mask R-CNN and ResNet101-XGBoost Hybrid Model

Abstract

In Road accidents are a major global issue, often resulting in loss of lives due to delayed response times. Traditional accident detection methods rely on manual reporting, which is slow and ineffective. This study presents an AI-driven system that detects and predicts accidents in real time, enabling faster emergency response and improved road safety.

The system processes CCTV footage using Convolutional Neural Networks (CNNs) to identify accident scenarios. Mask R-CNN is used for accident detection, while ResNet101 extracts key image features. Additionally, metadata such as vehicle speed, braking patterns, road conditions, and weather is analyzed using XGBoost to predict accident risks. If an accident is detected, an instant alert is sent to emergency responders.

To improve accuracy, the system has been trained on large datasets of accident and non-accident scenarios using techniques like data augmentation, transfer learning, and noise reduction. It achieves 97.67% accuracy in detection and 93.87% in prediction. The model easily integrates into smart city systems, enabling real-time accident detection and quicker emergency responses.