Hyperparameter Tuning of YOLOv8n for Real-Time Material Truck Detection

Abstract

The increasing number of material trucks on arterial roads has posed challenges for traffic surveillance and regulatory compliance. Traditional monitoring techniques that rely on manual observation are often ineffective and susceptible to irregularities, highlighting the need for automated real-time monitoring systems. This study proposes a lightweight object detection approach using YOLOv8n to improve real-time truck detection performance in traffic monitoring applications. A quantitative experimental methodology was employed by performing hyperparameter tuning through adjustments to the number of epochs, batch size, optimizer, and learning rate. The dataset was collected from real traffic environments using smartphone cameras and CCTV (TP-Link Tapo C320WS). A total of 36 experimental configurations were evaluated using Precision, Recall, F1-score, mAP@50, and mAP@50–95 metrics. Experimental results showed that the optimal configuration, consisting of 100 epochs, a batch size of 16, the Adam optimizer, and a learning rate of 0.001, achieved a mean Average Precision (mAP)@50 of 0.9302 and mAP@50–95 of 0.7226. Although the performance improvement over the baseline YOLOv8n model was relatively modest, repeated experiments demonstrated improved model stability and consistency after hyperparameter optimization. Real-time deployment on a local GPU achieved a stable processing speed of 14–23 Frames Per Second, with an average of 19 FPS, enabling real-time monitoring performance aligned with the camera input rate. The integrated system successfully combines object detection, tracking, and license plate recognition for practical traffic monitoring applications. However, smaller objects such as license plates remained more challenging to detect due to localization limitations under occlusion and low-light conditions.