Yolov8-Based Vehicle Detection and Traffic Density Estimation for Adaptive Traffic Signal Applications
TL;DR
The results support YOLOv8 as an efficient traffic-monitoring component, although density-classification accuracy and traffic-flow improvement require separate validation.
Abstract
Fixed-time traffic signals cannot respond to short-term changes in vehicle demand and may allocate green time inefficiently at urban intersections. This study developed and evaluated a YOLOv8-based vehicle-detection and traffic-density estimation prototype for adaptive signal applications. Two public traffic-intersection datasets supplied 3,300 images, of which 2,932 were retained after quality screening. Images were resized to 640 × 640 pixels, annotated with vehicle bounding boxes, and augmented through ±25% brightness and saturation adjustments. YOLOv8 was trained for 50 epochs using 70:30, 80:20, and 90:10 training-testing splits and was compared with Faster R-CNN under the 70:30 configuration. YOLOv8 achieved mAP@0.5 values of 0.989–0.992 and mAP@0.5:0.95 values of 0.962–0.965. Under the direct comparison, YOLOv8 obtained 0.989 mAP@0.5, 0.962 mAP@0.5:0.95, and 0.980 recall, while requiring 1.9 ms per image compared with 28.4 ms for Faster R-CNN. The detector was integrated into a web prototype that displayed vehicle counts, estimated density, and signal duration. The results support YOLOv8 as an efficient traffic-monitoring component, although density-classification accuracy and traffic-flow improvement require separate validation.