International Journal of Science, Engineering and TechnologyAuthors: Md Zahidul Islam Sany, Zhang Wubo
Abstract: Real-time object detection plays a foundational role in the deployment of Intelligent Transportation Systems (ITS) and modern smart city frameworks. To facilitate effective traffic management, automated intersection regulation, and active protection for vulnerable road users (VRUs), vision-based deep learning models must deliver high spatial localization accuracy without introducing heavy computational latencies. This research presents a comprehensive performance evaluation of the lightweight anchor-free YOLOv8 nano (YOLOv8n) architecture deployed on the CAVI-14 (Camera-based Vehicle Image) dataset. The network was trained over a rigorous 100-epoch horizon with an input resolution of 640 \times 640 pixels. The experimental findings reveal high performance capabilities across critical urban mobility profiles, yielding an overall bounding box precision of 93.7%, a recall of 96.1%, a mean Average Precision mAP50 of 96.9%, and a stringent mAP50-95 of 78.4%. Notably, highly specialized categories specifically ambulance and bicycle achieved targeted localized mAP metrics of 99.5%. Detailed examination of confidence landscapes revealed that the network reached its optimal macro F1-score of 0.95 at a confidence setting of 0.358, achieving total processing throughput of ~208 Frames Per Second (FPS). This systematic evaluation validates that highly compact, parameterized single-stage deep neural networks can successfully meet the low-latency, high-precision constraints mandatory for edge-level smart city infrastructure.
DOI: http://doi.org/10.5281/zenodo.20743880
