Counting Mixed Traffic Volumes at Motorcycle-Dominated Intersections by Using Computer Vision

This study addresses the lack of computer vision techniques for counting traffic at motorcycle-dominated intersections by developing an integrated framework. Three models are proposed: a detection model with a large visual dataset, a tracking model with a proposed reference point of bounding box and a suggested algorithm for cases of vehicle identity switches, and a counting model with algorithms for different junction movements. These models were applied to the case study of Vietnam. A large dataset comprising 52 footages recorded in daytime and nighttime conditions yields 1,195,691 labelled vehicles. The tracking model accurately reflects vehicle trajectories on road surface, while the counting model improves performance with a triple double-line method. The counting model achieve over 90% accuracy compared to manual counting in terms of total volume and each vehicle type. Therefore, transport planners and operators in Vietnam can draw on the findings of this research by applying the models to data collection, count traffic and monitor intersections. These models might be modified to other countries where motorcycles are dominant.