Abstract:
This thesis presents an intelligent traffic monitoring framework using YOLOv8 for realtime detection is proposed. This framework utilizes both traffic videos from UADETRAC and an emergency vehicle dataset to provide enhanced real time vehicle detection and densities of lane usage and estimation of emergency vehicle clearance time. The average pixels per second processed is 25 P/second for both datasets. All videos were preprocessed with standard preprocessing including normalization and augmentation of training data to aid in improving the robustness of the trained model. For the trained model, two YOLOv8 models were created, one for general vehicles and one specifically for emergency vehicle detection and training. These models performed well on both with respect to precision, recall, F1-score, mAP, and minimal cross-class confusion on both datasets. The vehicles detected in each frame will be counted for lane density estimation purposes; a phased and continuously time-based Gaussian model with time duration varied based on density of activity (density of traffic lanes) will visually express the congestion levels based on vehicle density and activity over time. This real time traffic control framework was simulated in a Python-based Adaptive Signal Control System demonstrating that the system would adjust the lengths of green light duration based on the traffic lane densities and the presence of EVs and emergency response vehicles. In addition to reducing rear-end and lane throughput delays, the system would support facilitating rapid clearance of ambulances and firetrucks. Overall, this proposed framework provides a practical, robust, transparent and reproducible methodology to combine real time object detection and high-level traffic management for the development of intelligent transportation systems.
