In the context of intelligent transportation systems real-time vehicle detection and tracking on highways present significant challenges due to the complexity of high-resolution imagery, varying lighting conditions and occlusions. Existing systems often struggle to balance computational efficiency with the ability to detect fine-grained vehicle attributes. This paper proposes FYNet, a novel architecture based on YOLOv5, designed for real-time vehicle localization and attribute identification. FYNet introduces a novel Path Aggregation Network to enhance multi-scale feature extraction, reduce computational overhead, and improve detection accuracy for objects of varying sizes, from license plates to long vehicles. With five outputs at different resolutions, FYNet achieves a robust detection across all object sizes, an inference speed of 16.3ms per 4K image (60 FPS) , and reduces computations to 0.6 GFLOPs. The StrongSORT method, which is based on DeepSORT, is used to track the vehicles and assign a single ID for each one. A simple yet effective strategy of separating front and rear views of vehicles into distinct classes also improved the model's mean average precision (mAP) by 1.8%. It also helped the model to recognize the characteristics of the vehicles with fewer parameters and higher accuracy. To evaluate the model, a private dataset, called IRVA, with more than 300K labels from 11 classes, is prepared. It has several new features compared to the existing industrial datasets in ITS. The FYNet outperforms the standard YOLOv5 in both inference speed and accuracy, achieving a 0.6% improvement in object recognition accuracy while maintaining real-time performance on 4K images.