A Synthetic Data Generation and Learning Framework Using Top-down View Transformation for Situation Awareness Enhancement in Unmanned Battlefield Systems

전장 상황 무인 시스템 상황인지 향상을 위한 Top-down 변환 기반 합성 데이터 생성 및 학습 기법

Military robots and unmanned systems require training data with diverse viewpoints for robust object detection and situational awareness. Top-down view images are difficult to acquire, resulting in limited viewpoint diversity. In this paper, we propose a pipeline that transforms ground view images into top-down view images while preserving structural consistency. The framework employs FLUX.2 with Low-Rank Adaptation (LoRA) for parameter-efficient viewpoint adaptation. Instead of updating the entire network, we apply LoRA only to the Cross-Attention layers, which are critical for viewpoint conditioning, while freezing the pretrained network. We demonstrate that this approach effectively performs viewpoint transformation in data-limited military domains using only a small-scale training dataset. In addition, we employ conditional generation with ground view inputs, prompts, and target images. Our approach adapts a large pretrained model using only small datasets to generate structurally consistent top-down synthetic images. Structural consistency analysis based on LightGlue correspondences and CLIPPER-based geometric filtering yields a maximum inlier ratio of 52.7%. Finally, training with the synthetic data improves the mean Average Precision (mAP) by 6.5 percentage points over the original-data-only baseline.