Advanced UAV-based Multi-Vehicle Monitoring using Deep learner and Vision Transformer

Urbanization and the increasing number of vehicles on roads have overwhelmed traditional traffic management systems, leading to heightened congestion, accident rates, and environmental degradation. To address these challenges, this paper proposes a novel deep learning-based framework for comprehensive vehicle detection, tracking, and classification in aerial imagery, providing a scalable solution for modern traffic surveillance. The framework employs Contrast Limited Adaptive Histogram Equalization (CLAHE) for enhanced image preprocessing, followed by precise segmentation using the SegFormer model. Vehicle detection is handled by the advanced YOLOv10 architecture, which is optimized for aerial views and excels in detecting vehicles in complex, occluded environments. The proposed system integrates TrackFormer for robust multiobject tracking, DBSCAN for accurate vehicle counting, and a hybrid feature extraction pipeline utilizing CNNs, SIFT, and BRISK for detailed feature extraction. Classification is further enhanced by CrossViT, a cross-attention vision transformer that improves accuracy across varying vehicle scales, while Graph Neural Networks (GNNs) are employed to predict vehicle trajectories. Experimental evaluation of our proposed framework, conducted on the VAID dataset, revealed a vehicle detection efficiency of 95%, a tracking capacity of 91%, and a mean classification accuracy of 93%. These findings affirm the effectiveness of our framework in addressing the complexities of large-scale aerial monitoring, positioning it as an innovative solution for enhancing intelligent transportation systems.