YOLOV10 vs. YOLOV8: Performance Improvements for Vehicle Detection at Multilane Roundabouts

This study compares YOLOv8 and YOLOv10 for vehicle detection in multilane roundabouts, a complex environment characterized by frequent occlusions and non-linear movement patterns. A large UAV-based dataset was collected from multiple roundabouts, capturing diverse geometric configurations and congestion levels. The dataset was preprocessed, annotated, and used to train and evaluate both algorithms under identical conditions. The results reveal that YOLOv10 models achieve higher recall across all vehicle categories, making them more effective for complete vehicle detection, including those with occlusions. In contrast, YOLOv8 models maintain slightly higher precision, reducing false positives, which is advantageous in applications prioritizing classification accuracy. Larger YOLOv10 models also demonstrate lower inference latency, thereby improving the feasibility of real-time deployment. However, YOLOv8 offers faster training times and is more memory-efficient in smaller configurations. Therefore, YOLOv8 is suitable for resource-limited environments that require high precision, while YOLOv10 excels at recall-oriented tasks. These findings contribute to the advancement of vehicle detection technologies, supporting the development of more effective and scalable intelligent transportation systems.