Real-Time Vehicle Classification and Directional Counting at Urban Roads Using YOLOv11 Under Heterogeneous Traffic Conditions

Urban roads in developing countries face highly heterogeneous traffic, making manual data collection labour-intensive and error-prone. This study presents a deep learning-based approach for real-time vehicle classification and directional counting using the YOLOv11 framework. A custom dataset of 5,200 locally collected images, encompassing 11 vehicle classes defined by the Road Development Authority of Sri Lanka, was used to train two model variants: YOLOv11 medium and YOLOv11 small. To enable directional flow estimation, the system implements a line-pass-based tracking algorithm that distinguishes between up- and down-traffic streams. The system demonstrated high accuracy for common classes, such as cars (Upstream 98.8% and Downstream 96.8%), motorcycles (Upstream 88.8% and Downstream 97.6%), and three-wheelers (Upstream 85.8% and Downstream 98.4%). However, performance variations in other categories highlight the challenges of occlusion and visual ambiguity in mixed-traffic environments. The findings demonstrate that YOLOv11 models trained on locally representative datasets provide a reliable, scalable, and cost-effective alternative to manual survey methods. This approach provides a disaggregated data source to support traffic signal optimisation, intersection analysis, and urban transport planning under complex heterogeneous traffic conditions.