Vehicle tracking for mobility applications

The goal of this project was to evaluate various object detection and tracking models to determine the best combination for real-time vehicle monitoring in urban settings, with a focus on parking spot tracking. This study examined multiple algorithms, including different sizes of YOLOv8 for detection and OC-SORT and StrongSORT for tracking. Both public and private datasets were used to provide a more complete evaluation. In the experiments, YOLOv8 emerged as an effective detection model, with varying trade-offs between speed and accuracy depending on the model size. YOLOv8n, the smaller version, provided higher frames per second (FPS) and lower computational overhead, making it suitable for real-time applications. In contrast, YOLOv8l, the larger version, exhibited greater detection accuracy with longer inference times, suggesting its potential use in applications where precision is outstanding. For tracking, the study compared OC-SORT and StrongSORT. StrongSORT consistently demonstrated better tracking accuracy but with a slightly lower FPS. OC-SORT, on the other hand, achieved higher FPS, making it more suitable for real-time scenarios. Additionally, StrongSORT proved more robust against occlusions and non-linear object motion, thanks to its advanced modules, while OC-SORT excelled in simpler tracking tasks. Based on these results, the ideal solution for a traffic monitoring application focusing on parking spot tracking would be combining YOLOv8n for fast and efficient detection with StrongSORT for robust tracking. This combination offers a balance of speed, accuracy, and computational efficiency, allowing for real-time monitoring while ensuring precise tracking. A key aspect for accurate vehicle tracking is high-resolution video, with 4K being the optimal choice for detailed detection and license plate recognition. The experiments also suggested that for applications requiring faster real-time online tracking, YOLOv8n coupled with OC-SORT might be a viable soluti