Design and implementation of VTAS-Net for multi-vehicle detection and tracking using YOLOv8 and deep SORT

The proposed paper is an enhanced deep learning-based system, VTAS-Net (Vehicle Tracking and Analytics System), comprising of Deep SORT algorithm integrated with YOLOv8m to detect vehicles in real-time and track multiple objects. The proposed framework allows proper and precise tracking of vehicles in complex traffic conditions using the anchor-free architecture of the YOLOv8m, the ability to fuse features adaptively and the attention-based representation together with the identity-conserving property of Deep SORT. The COCO dataset, the Berkeley Deep Drive (BDD100K) dataset, and a bespoke dataset with more than 5000 annotated frames from various highway and urban scenarios are used to assess the system. Experimentally, VTAS-Net is better than existing models, such as YOLOv4, YOLOv5m, YOLOv6, and YOLOv7, with a mAP50 of 78.5 and a vehicle counting accuracy of 94.2. Additionally, the integration of Deep SORT lowers identification switch rates by 28% and false positives by about 30%, guaranteeing steady and reliable multi-vehicle tracking in high traffic situations. The proposed structure can be implemented to smart transportation systems because the inference rate of the structure is up to 5.4 ms/frame, and real-time processing can exceed 180 FPS on the platform with a GPU. Even with its excellent performance, it still has trouble identifying tiny and highly obscured cars. Future studies will focus on optimisation of edge-device deployment, bettering of occlusion management, and prediction analytics to manage intelligent traffic. Overall, VTAS-Net provides a scalable, efficient, and practical solution for real-time traffic monitoring and advanced smart mobility applications.