Hyperparameter Tuning Strategy for YOLOv8n in Real-Time Post-Accident Traffic Monitoring

Abstract

Traffic accidents continue to provide a considerable difficulty in contemporary transportation systems, frequently leading to vehicle damage and heightened risks for pedestrians on streets. Precise and instantaneous identification of post-accident scenarios is thus crucial for facilitating swift response and sophisticated traffic management. This research introduces a streamlined object detection methodology utilizing YOLOv8n to recognize six essential traffic-related categories: bus, automobile, damaged vehicle, motorbike, pedestrian, and truck. The main aim is to examine the impact of hyperparameter modification on detection efficacy, specifically in recognizing damaged automobiles as signs of post-accident situations. Twelve model configurations were created by systematically altering three hyperparameters: learning rate (0.01, 0.001, and 0.0001), batch size (32 and 64), and optimizer type (Adam and MuSGD). All models underwent training for 200 epochs with a dataset derived from actual traffic situations, augmented by techniques such as grayscale transformation, blurring, and rotation. The performance evaluation utilized precision, recall, F1-score, mAP50, and mAP50:95. The findings indicate that hyperparameter selection substantially influences convergence stability and detection accuracy. The optimal model attained a mAP50 of 0.905 and a mAP50:95 of 0.751, utilizing a learning rate of 0.01, a batch size of 64, and the Adam optimizer. Moreover, substantial items like cars, buses, and trucks were identified with high precision, whereas damaged vehicles and pedestrians necessitated more meticulous calibration due to increased visual variability.The findings indicate that optimized lightweight models can attain competitive performance, rendering them appropriate for real-time intelligent traffic monitoring applications.