無人機在城市環境煙霧和樹葉遮蔽下的坦克偵測問題

Tank Detection with Occlusions of Smoke and Foliage for UAV in Urban Environments

現有的無人機影像資料集，例如VisDrone，提供了豐富的空拍城市影像，但不包含坦克物件類別；而現有的坦克影像資料集通常缺乏城市影像背景和城市環境煙霧樹葉遮蔽。為了彌補這一不足，本文提出了一個合成無人機空拍坦克影像與城市影像的生成框架，將城市影像與坦克影像結合，並自動產生YOLO格式的標註，並在隨後產生各種黑煙、白煙和樹葉遮擋，以進行遮蔽穩健性評估。YOLOv10-S在測試集上取得了99.0%的mAP 50，包含89.7%的精確率和98.8%的召回率。在黑煙和白煙條件下分別取得了96.7%和96.6%的mAP 50。在80%樹葉遮蔽下，取得了88.8%的mAP 50、73.6%的精確率和88.5%的召回率。

Existing UAV (Unmanned Aerial Vehicle) image datasets, such as VisDrone, provide diverse urban aerial imagery but do not include a tank category, whereas available tank image datasets generally lack urban UAV backgrounds and controlled occlusion conditions. To address this gap, this paper proposes a synthetic urban tank image generation framework that combines urban UAV background images with tank foreground images and automatically produces YOLO-format annotations. Black-smoke, white-smoke, and foliage occlusion conditions are then generated for controlled robustness evaluation. YOLOv10-S achieved 99.0% mAP 50, 89.7% Precision, and 98.8% Recall on the test-set. On the occlusion validation subsets, matched training achieved 96.7% and 96.6% mAP 50 under black- and white-smoke conditions, respectively. Under 80% foliage occlusion, matched training achieved 88.8% mAP 50, 73.6% Precision, and 88.5% Recall.