YOLOv8 vs. YOLOv9: Safety Helmet Detection Performance

We evaluated the performance of YOLOv8 and YOLOv9 for helmet detection in construction worker images using the mAP metric, training duration (hours), and the number of epochs (25, 50, and 100). First, we collect images online with several conditions, such as various image sizes and various image quality. Then, we annotate the photos, defining which part of the image contains the heads of people wearing helmets and those not wearing helmets. To overcome limited data, we applied data augmentation by rotating and flipping the images. Moreover, the data for training increased. Then, we divided the data into three parts: training set, validation set, and testing set. We applied YOLOv8 and YOLOv9 frameworks to generate an AI model based on the training set and validation set, and last, we evaluated the model using the testing set. As a result, YOLOv8 has a shorter training time, but at epoch 100, its performance degrades slightly. The mAP values of YOLOv8 for epochs 25, 50, and 100 are 0.973, 0.974, and 0.956. Training 100 epochs takes 0.666 hours, while YOLOv9 takes 1.461 hours. Despite taking longer, YOLOv9 provides more consistent results with mAP scores of 0.974, 0.972, and 0.976. This result indicates a more robust learning capability. The choice of model depends on the need between training speed and precision level.