UAV-based Road Traffic Monitoring via FCN Segmentation and Deepsort for Smart Cities

Efficient vehicle detection and tracking are crucial for modern traffic management, yet traditional systems are limited by fixed infrastructure. Aerial imagery, offering extensive coverage and flexibility, presents a transformative alternative. This paper introduces an advanced vehicle detection and tracking framework within aerial image sequences, combining cutting-edge image processing and deep learning techniques. Our approach utilizes a Histogram Equalizationbased preprocessing pipeline to enhance image contrast and reduce noise. A Fully Convolutional Network (FCN) is employed for precise semantic segmentation, followed by the integration of the Deep SORT algorithm for dynamic detection. The SuperGlue algorithm ensures robust feature matching across frames, preserving vehicle identity. RetinaNet is used for accurate vehicle counting in complex environments, while the Kalman filter framework guarantees high-precision tracking and trajectory consistency. Experimental results on the VAID dataset show a detection accuracy of $\mathbf{9 5 \%}$ and tracking accuracy of 91%, highlighting the methodology’s potential to revolutionize real-time traffic management. This integrated solution sets a new standard for vehicle tracking in aerial imagery and offers significant implications for future smart city application.