Lightweight YOLO Models for Real-Time Multi-Vehicle Detection

This study presents a comparative evaluation of three lightweight YOLO architectures: YOLOv5n, YOLOv8n, and YOLOv11n, for multi-class vehicle detection using CCTV imagery captured under dense traffic conditions in Semarang, Indonesia. The models were tested on their ability to detect four types of vehicles, including motorcycle, car, bus, and truck. To enhance generalization across different lighting conditions, image qualities, and environmental noise, six data augmentation techniques were applied during training. These included Blur, Brightness Adjustment, Color Jitter, Noise Injection, Scaling, and Zoom In. Among these, the Blur technique yielded the most significant improvement in detection accuracy. YOLOv8n with Blur augmentation achieved the best performance with a precision of 0.875, recall of 0.655, [email protected] of 0.756, and [email protected]:0.95 of 0.467. Class-wise analysis showed that buses and trucks were easier to detect due to their larger size and distinct features, while motorcycles were the most difficult due to their smaller dimensions and visual similarity to other objects. Training curves demonstrated consistent decreases in loss values and progressive improvements in evaluation metrics across 60 epochs. These findings emphasize the importance of selecting appropriate model architecture and augmentation strategies to improve object detection performance, particularly in real-time and resource-limited applications. YOLOv8n with Blur augmentation proved to be the most effective configuration in this study.