Development and Optimization of YOLOv12 for Autonomous Vehicle Navigation Systems

Abstrak

This study develops and enhances a YOLOv12-based object detection model for autonomous vehicle perception on Indonesian highways, addressing limitations of earlier research that lacked realistic traffic scenarios and field validation. The Roboflow dataset contains 29 object classes, including vehicles, pedestrians, and traffic signs, with existing annotations. Preprocessing included data quality assessment, image resizing, dataset split validation, annotation format conversion, and data augmentation to improve training performance. Eight training configurations were evaluated by varying learning rate, batch size, and optimizer. Initial comparisons showed YOLOv12 significantly outperformed SSD, achieving mAP50 of 0.978 and mAP50–95 of 0.831, compared to SSD’s 0.816 and 0.639. SGD consistently provided more stable and accurate performance than Adam. The best model used SGD with a learning rate of 0.001 and batch size of 16, achieving precision of 0.952, recall of 0.955, mAP50 of 0.974, and mAP50–95 of 0.834. Field testing confirmed strong detection of pedestrians and traffic signs, although challenges remained with small and overlapping objects. Future work should improve small-object detection, expand dataset diversity, and explore advanced architectures or hybrid optimization strategies. These findings support YOLOv12 as a reliable foundation for safer, more efficient self-driving perception systems tailored to Indonesia’s complex road environments in real conditions.