Optimization model of small target detection for vehicle extraction in virtual simulation scenarios

In modern virtual simulations for autonomous driving, traditional scene generation methods based on expert-driven enumeration are inadequate to meet the evolving requirements of advanced simulation environments. Therefore, research into advanced virtual scene generation techniques has become essential. To address this challenge, we enhance small target detection models for vehicles in virtual scenes using the YOLOv10n framework by optimizing its network architecture to improve detection accuracy. This study integrates literature reviews and empirical experiments to evaluate and refine vehicle small target detection and model precision. We design an ablation study to systematically assess various modifications and determine the optimal configuration. The final enhanced model demonstrates a 5.19% improvement in accuracy relative to the original model.