TK1113 : Real-Time Vehicle Detection and Recognition Using an Enhanced YOLO Network and Their Tracking in Multi-Lane Highways
Thesis > Central Library of Shahrood University > Electrical Engineering > PhD > 2026
Authors:
[Author], Hossein Khosravi[Supervisor]
Abstarct: Abstract Real-time detection and recognition of vehicles in intelligent transportation systems, particularly in multi-lane highways, face significant challenges. These include processing ultra-high-resolution 4K images, handling a large number of objects within a single frxame, managing the wide diversity of vehicle types, detecting objects ranging from very small to very large, coping with occlusions in congested environments, and maintaining robustness against varying illumination conditions and diverse camera angles. Existing systems, primarily baxsed on standard YOLO architectures, often lack sufficient capability to extract fine-grained vehicle features and to ensure stable tracking on high-resolution 4K imagery in multi-lane highway scenarios. To address these challenges, this dissertation introduces a novel architecture termed Flexible YOLO Network (FYNet), developed upon YOLOv5. The proposed model incorporates several optimizations, including intelligent reduction of filter width, increased network depth, replacement of the conventional PAN block with the innovative PANSum block, and integration of a Head-5D structure in the detection head. These enhancements enable a balanced trade-off between accuracy and computational efficiency. With five output branches at different scales, FYNet effectively detects objects of varying sizes—from license plates to heavy trucks—while reducing computational cost from 1.2 GFLOPs to 0.6 GFLOPs. For vehicle tracking, the StrongSORT algorithm is employed to assign unique identifiers to each vehicle, ensuring reliable tracking under occlusion. Furthermore, this research introduces a novel classification strategy that distinguishes between front and rear views of vehicles, resulting in an improvement of mAP by 1.8%. The proposed model is evaluated on a newly constructed private dataset, IRVA, comprising over 300,000 annotations across 11 classes, collected under diverse lighting conditions, both day and night, and from multiple viewing angles. Experimental results demonstrate that FYNet achieves an mAP of 94% on 4K images with an inference time of 16.33 ms per frxame (equivalent to 60 FPS), outperforming the standard YOLOv5 in both detection accuracy and inference speed. The novelty of this dissertation lies in the introduction of the FYNet architecture for real-time vehicle detection and classification, the design of the PANSum block to enhance feature representation while reducing computational load, the utilization of multi-scale output branches to improve detection across diverse object sizes, the integration of StrongSORT for robust real-time tracking, and the development of the IRVA dataset with unique characteristics for model evaluation. Collectively, these contributions culminate in a comprehensive monitoring and control ering a six-lane highway using only a single 4K camera and a single processor, offering a simple, cost-effective, and reliable solution for real-time applications in intelligent transportation systems.
Keywords:
#Keywords: YOLOv5; FYNet; Real-time localization; DeepSORT; Intelligent Transportation System Keeping place: Central Library of Shahrood University
Visitor: