Robust Vehicle Detection and Tracking in Aerial Videos in Expressway Merge Areas
Yu-Wei Fang, Bi-Jiang Tian, Xing-Tong Chen, Peng Li, Wenchen Yang, Feng Zhu · Journal of Advanced Transportation · 2026
Vehicle trajectory extraction from unmanned aerial vehicle (UAV) aerial videos offers valuable data for traffic safety analysis, especially in high‐risk areas such as expressway merge areas. However, the unique characteristics of aerial footage, such as small object sizes, limited distinguishing features, and frequent tracking ID switches, pose significant challenges to accurate detection and tracking. This study proposes a robust framework to address these challenges and enhance vehicle trajectory reconstruction. The YOLOv5 detection model is integrated with a dedicated small‐target prediction head and attention mechanisms to improve detection precision under complex backgrounds. To boost tracking accuracy and efficiency, the DeepSORT algorithm is modified by replacing its feature extraction backbone with a lightweight MobileNetV2 network. The resulting trajectories are further refined through filtering and transformed into ground coordinates. A tree‐based reconstruction algorithm is applied to enhance temporal continuity. Experimental results show notable improvements: vehicle detection accuracy increased by 8.18%, 8.96%, and 12.62% at aerial altitudes of 200, 250, and 350 m, respectively; multiobject tracking accuracy improved by 6.7%, 8.27%, and 10.9% at the same altitudes. These outcomes demonstrate the effectiveness of the proposed framework for UAV‐based traffic analysis in complex environments such as expressway merge areas.