Abstract:
Urbanization and increasing traffic volumes have intensified congestion,
highlighting the need for intelligent transportation solutions. Traditional traffic data
collection methods are often inaccurate and inefficient. This thesis proposes an
advanced traffic analysis system using YOLOv8 for object detection and DeepSORT
for tracking to estimate traffic parameters like volume, density, and speed from video
inputs. A dataset of 20,000 images across seven vehicle classes (Car, Bike, Bus,
Truck, Hiace/Van, Tractor, and Rickshaw) was annotated using Roboflow. The
YOLOv8 model demonstrated high precision and recall, with a mAP@0.5 of 0.88
and a mAP@0.5:0.95 of 0.64. DeepSORT effectively tracked moving objects,
maintaining unique ID’s across frames for accurate traffic parameter estimation. The
developed API integrates detection and tracking data from video streams, providing
precise traffic volume, density, and speed estimates. Tests on video clips yielded
accurate traffic volume estimates of 992, 6,397, and 2,879 vehicles per hour for the
G10-F10 and Srinagar Highway F-9 intersection. It also determined traffic densities
of 21.94, 19.81, and 31.25 pcu/mile/lane and speeds of 27.46, 53.81, and 31.82 mph
for the same locations. This research showcases the potential of real-time YOLOv8
and DeepSORT for traffic analysis. Future improvements could include expanding
the dataset, fine-tuning the model, integrating additional data sources, and applying
the system to real-world traffic monitoring. The proposed system offers valuable
insights for traffic control and urban planning, providing a foundation for enhancing
transportation infrastructure's safety and efficiency.
