Scholarlink

<p>Aerial imagery is a vital tool in intelligent traffic management and urban planning. However, the presence of dynamic objects, such as vehicles and pedestrians, often obstructs road signs and traffic markings, leading to a decrease in the accuracy of image analysis. While traditional methods can partially restore the background, they are plagued by high computational costs, blurred details, and background misalignment when dealing with large-scale aerial data. In response to these challenges, this study introduces a novel background reconstruction method that integrates global alignment, object recognition, and weighted averaging. The process begins with RANSAC with homography, aligning multiple images and correcting spatial errors caused by UAV shake or viewpoint shifts. Subsequently, in the aligned image sequence, dynamic objects are identified through a combination of depth estimation, frame differencing, and the YOLOv12 model, resulting in precise dynamic masks. At this stage, the detection model achieves a Precision of 0.917, a Recall of 0.886, an F1-score of 0.901, and an AP of 0.946, showcasing its robust performance. Finally, the dynamic masks are used in a weighted background averaging process to eliminate dynamic regions and restore a stable, transparent background. Experimental results demonstrate that the proposed method surpasses traditional averaging and generative models (such as GANs and Diffusion) in preserving traffic details like crosswalks, lane markings, and directional indicators. This approach could apply to areas such as intelligent traffic monitoring, UAV image analysis, and urban planning, indicating its potential relevance in practical scenarios. The project of this work is made publicly available at <a href="https://github.com/seannnnnn1017/dynamic-object-removal-aerial">https://github.com/seannnnnn1017/dynamic-object-removal-aerial</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic object removal and background reconstruction in aerial images via global alignment and YOLOv12

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Chi Hung Wang,
  • Yu Siang Siang,
  • Xiang Shun Yang,
  • Jun Jie Yen,
  • Wei Ren Chen

摘要

Aerial imagery is a vital tool in intelligent traffic management and urban planning. However, the presence of dynamic objects, such as vehicles and pedestrians, often obstructs road signs and traffic markings, leading to a decrease in the accuracy of image analysis. While traditional methods can partially restore the background, they are plagued by high computational costs, blurred details, and background misalignment when dealing with large-scale aerial data. In response to these challenges, this study introduces a novel background reconstruction method that integrates global alignment, object recognition, and weighted averaging. The process begins with RANSAC with homography, aligning multiple images and correcting spatial errors caused by UAV shake or viewpoint shifts. Subsequently, in the aligned image sequence, dynamic objects are identified through a combination of depth estimation, frame differencing, and the YOLOv12 model, resulting in precise dynamic masks. At this stage, the detection model achieves a Precision of 0.917, a Recall of 0.886, an F1-score of 0.901, and an AP of 0.946, showcasing its robust performance. Finally, the dynamic masks are used in a weighted background averaging process to eliminate dynamic regions and restore a stable, transparent background. Experimental results demonstrate that the proposed method surpasses traditional averaging and generative models (such as GANs and Diffusion) in preserving traffic details like crosswalks, lane markings, and directional indicators. This approach could apply to areas such as intelligent traffic monitoring, UAV image analysis, and urban planning, indicating its potential relevance in practical scenarios. The project of this work is made publicly available at https://github.com/seannnnnn1017/dynamic-object-removal-aerial.