2026 ASEE Annual Conference & Exposition

The paper investigates the application of AI in enhancing safety in construction zones.
Urban construction activities introduce significant volatility into traffic networks, creating dynamic, high-risk conditions that often overwhelm conventional traffic management systems. Fixed-time or actuated signal plans lack the responsiveness to adapt to irregular traffic flows, unpredictable lane closures, and evolving safety risks that characterize construction zones. Addressing these limitations, the paper presents a comprehensive AI-driven traffic management framework that integrates advanced data science methodologies, machine learning models, and real-time computing architectures to enhance efficiency and safety in construction-impacted urban corridors.

At its core, the proposed system fuses multimodal data streams from IoT sensors (inductive loops, radar units), high-resolution video analytics, and connected-vehicle telemetry into a unified, high-frequency dataset for intelligent decision-making. The data pipeline supports specialized AI models for distinct operational objectives. For environmental perception, a YOLO-based object detection network identifies and tracks vehicles, cyclists, and pedestrians. At the same time, a lightweight ResNet-based Convolutional Neural Network performs anomaly detection to recognize stalled vehicles, unauthorized pedestrian movements, and other safety hazards in real time. Predictive analytics through XGBoost regressors and Long Short-Term Memory networks forecast short-term traffic volumes and queue lengths by leveraging temporal features such as weather, time of day, and construction activity patterns.

The central innovation lies in an adaptive signal control mechanism powered by deep reinforcement learning. These RL agents are trained within a high-fidelity SUMO-based digital twin simulation, modeling stochastic traffic arrivals and dynamic construction layouts to ensure robust, transferable control policies. The agents continuously learn optimal strategies for adjusting signal phases, cycle lengths, and offsets through a reward function that jointly minimizes traffic delay and safety risk. Pre-trained policies are fine-tuned to bridge simulation and deployment using transfer learning on field-collected data, improving real-world adaptability.

The system operates within a hybrid cloud edge computing architecture for real-time responsiveness and scalability. At the tactical level, edge inference nodes embedded in traffic cabinets execute low-latency analytics and control actions, maintaining sub-100 millisecond response times. At the strategic level, the cloud orchestration layer oversees data storage, model retraining, and long-term optimization. Complementing the control framework, a Gradient-Boosted Trees based safety module computes real-time risk scores derived from traffic variance, gap-time fluctuations, and obstruction flags. These scores trigger proactive interventions such as dynamic lane reassignment, adaptive all-red intervals, or driver alerts via Variable Message Signs to reduce worker and driver hazards.

Comprehensive simulation experiments benchmark the proposed framework against traditional fixed-timing and actuated systems under randomized, high-stress construction scenarios. The results demonstrate compelling performance gains: up to 30% higher intersection throughput, 25% shorter travel times, and 20% fewer near-miss events. These outcomes underscore the potential of AI-based adaptive systems to transform urban traffic management from reactive control to predictive, data-driven optimization.