DGTS-YOLOv8n: a lightweight vehicle detection algorithm based on YOLOv8n

To enable real-time vehicle detection in computationally constrained scenarios, we propose DGTS-YOLOv8n, an efficient model based on an enhanced YOLOv8n architecture. Key improvements include: (1) replacing C2f with GhostC2f modules and adopting GhostConv in backbone/neck networks to reduce computational costs; (2) integrating triplet attention to enhance feature representation; and (3) utilizing soft-non-maximum suppression (NMS) for improved post-processing. Experiments on the UA-DETRAC dataset demonstrate that DGTS-YOLOv8n achieves a 37% reduction in FLOPs and 48% fewer parameters while maintaining competitive accuracy (only 0.2% mAP@.5:.95 degradation). On the Jetson Xavier NX edge computing platform, it achieves real-time inference at 68.6 FPS, making it suitable for deployment in embedded artificial intelligence devices.