Analysis of Hyperparameter Influence on YOLOv8 Performance in Multimodal Image Sets

Object detection in images is essential for applications such as security, industrial monitoring, and autonomous vehicles. Deep learning models like YOLO (You Only Look Once) are widely used for their accuracy in identifying and localizing objects. This study investigates the optimization of YOLOv8 by tuning hyperparameters, using a labeled dataset that includes multimodal image sets captured by multisensors. These images exhibit diversity in scenarios, including visible and thermal images of cars, motorcycles, and trucks captured by remotely piloted vehicles on rural and paved roads. The capture conditions on rural roads add significant challenges to data acquisition, such as terrain variability and weather conditions. We evaluate the impact of variations in dropout rate, batch size, and single-class configuration (single_cls) on performance metrics: precision, recall, mAP50, and mAP50-95. The results provide valuable insights for optimizing YOLOv8 in different object detection scenarios.