<![CDATA[Intelligent transportation systems required rapid and highly accurate object detection algorithms to function effectively in real-time environments. Although the You Only Look Once (YOLO) architecture evolved significantly to address these demands, comprehensive performance comparisons between its latest iterations remained sparse in the context of urban traffic. Selecting the optimal model for specific deployment constraints remained a critical challenge for developers. To address this gap, this research comparatively evaluated the performance of YOLOv8 and the newly released YOLOv11 specifically for vehicle detection tasks. The primary contribution was providing a rigorous, empirical head-to-head comparison to guide model deployment based on specific accuracy needs and hardware constraints. The methodology employed a standardized public dataset of traffic imagery containing various vehicle classes. Both nano variants, namely YOLOv8n and YOLOv11n, were trained from scratch under identical hardware environments and hyperparameter settings. Specifically, the training utilized the AdamW optimizer for fifteen epochs with a batch size of sixteen. Performance was quantitatively measured using Precision, Recall, mean Average Precision (mAP50 and mAP50-95), and computational resource allocation in terms of parameter size. The experimental results indicated a nuanced performance dynamic. YOLOv11n demonstrated a vastly superior Precision score of 0.784 compared to 0.672 for YOLOv8n, signifying its enhanced reliability with fewer false alarms due to refined spatial attention modules. Conversely, YOLOv8n outperformed YOLOv11n in Recall (0.695 versus 0.662) and mAP50 (0.746 versus 0.740). This suggested that YOLOv8n achieved faster learning convergence under constrained training conditions. Both models maintained real-time processing capabilities suitable for edge devices. In conclusion, YOLOv11n was recommended for high-accuracy traffic surveillance systems prioritizing precision, whereas YOLOv8n remained a highly viable architecture for hardware-constrained applications that prioritized high recall and aggressive object detection.]]>
