Abstract

Qatar’s rapid urbanization and population growth have led to a significant increase in vehicle ownership and traffic accidents, creating challenges for public safety, emergency response, and urban planning. This study proposes TrafficTransformer, a deep learning model designed to predict traffic accident occurrences across 98 zones using traffic data collected from January 2017 to October 2023. The dataset, sourced from police traffic reports, includes hourly accident logs with features such as time of day, weather conditions, and spatial area codes. TrafficTransformer leverages self-attention and multi-head attention mechanisms to capture dynamic spatiotemporal patterns more effectively than traditional CNN and LSTM models. Model performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE), based on a 75%-25% training-testing split. Results demonstrate substantial improvements over baseline models. This research provides actionable insights for intelligent traffic management, targeted interventions, and resource optimization in emergency response systems.

Statistics