Unmanned aerial vehicles have revolutionized traffic monitoring systems by providing a unique top-view perspective crucial for accurately detecting and tracking vehicles in complex traffic scenarios. However, the efficiency of deep learning (DL) models in these environments can be significantly influenced by factors such as roundabout geometry and traffic congestion. Despite advancements in artificial intelligence-based detection and tracking algorithms, a notable gap remains in understanding how these factors specifically affect performance, especially in multilane roundabouts, where vehicle interactions are more complex. To fill this gap, this research investigates their effect on the performance of DL algorithms, You Only Look Once v8 and Deep Simple Online and Real Time Tracking, for vehicle detection and tracking, respectively. The results revealed that the geometric shape slightly affects vehicle detection accuracy, with elliptical and stadium-shaped roundabouts exhibiting higher precision and recall than round-shaped ones. However, the geometric shape has a negligible effect on tracking performance across shapes. Increased congestion significantly decreases detection precision and recall because of higher vehicle overlap, making it challenging to distinguish individual vehicles. Severely congested conditions present the lowest detection accuracy, highlighting difficulties in high traffic environments. An inverse relationship between congestion and testing speed was observed, with severe congestion leading to slower processing. Moderate congestion levels provide optimal tracking, and extreme congestion challenges accuracy, underscoring the need for adaptive strategies. Both factors independently affect DL performance; traffic congestion exerts a more significant influence. This research concludes that congestion affects detection and tracking performance more than geometry. Future research should focus on developing adaptive strategies tailored to specific traffic conditions.
Hamad et al. (Thu,) studied this question.