Scholarlink

<p>There is an urgent need for accurate and efficient methods to detect and count animals in the wild to support biodiversity conservation and ecosystem management. Traditional approaches to wildlife monitoring are often labor-intensive and time-consuming, while remote sensing technologies offer a scalable solution but usually struggle with feature extraction in complex environments. Challenges such as camouflage, background clutter, and the small size of animals lead to limited precision and accuracy. To address these challenges, this paper presents WL-YOLOv11, an enhanced version of the YOLOv11 model optimized for detection in complex environments like wildlife. The WL-YOLOv11 model incorporates a Coordinate Attention module in place of the Cross-Stage Partial with Self-Attention (C2PSA) module, replaces the C3K2 module with the Cross-Stage Partial Feature Fusion (C2f) module, and adopts Ghost Convolutions to improve both accuracy and computational efficiency of the model. The model is trained on a novel dataset that we extracted and annotated from real-life drone-captured videos of animals recorded at altitudes ranging from 20 to 150&#xa0;m. The dataset includes over 9k images of seven species: zebra, giraffe, wildebeest, elephant, impala, lechwe, and tsessebe. The images are challenging due to environmental camouflage, occlusion, and shadowing, especially the images taken from high altitudes. In addition to animal detection, we also provide a counting mechanism that accurately estimates the number of animals present in each image. Through the conducted experiments, the proposed model achieved a mean average precision (mAP) of 0.782 and a counting mean absolute error (MAE) of 3.73, outperforming the original YOLOv11 model by 3.3% in mAP and reducing the MAE from 3.95. The results highlight the model’s effectiveness in high-precision detection and counting of wildlife in complex natural environments, offering a scalable and real-time solution for conservation and ecological monitoring applications.</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

WL-YOLOv11: enhanced YOLOv11 for wildlife detection and counting in complex environments

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Amr Elkholy,
  • Amany M. Sarhan,
  • Aboul Ella Hassanien

摘要

There is an urgent need for accurate and efficient methods to detect and count animals in the wild to support biodiversity conservation and ecosystem management. Traditional approaches to wildlife monitoring are often labor-intensive and time-consuming, while remote sensing technologies offer a scalable solution but usually struggle with feature extraction in complex environments. Challenges such as camouflage, background clutter, and the small size of animals lead to limited precision and accuracy. To address these challenges, this paper presents WL-YOLOv11, an enhanced version of the YOLOv11 model optimized for detection in complex environments like wildlife. The WL-YOLOv11 model incorporates a Coordinate Attention module in place of the Cross-Stage Partial with Self-Attention (C2PSA) module, replaces the C3K2 module with the Cross-Stage Partial Feature Fusion (C2f) module, and adopts Ghost Convolutions to improve both accuracy and computational efficiency of the model. The model is trained on a novel dataset that we extracted and annotated from real-life drone-captured videos of animals recorded at altitudes ranging from 20 to 150 m. The dataset includes over 9k images of seven species: zebra, giraffe, wildebeest, elephant, impala, lechwe, and tsessebe. The images are challenging due to environmental camouflage, occlusion, and shadowing, especially the images taken from high altitudes. In addition to animal detection, we also provide a counting mechanism that accurately estimates the number of animals present in each image. Through the conducted experiments, the proposed model achieved a mean average precision (mAP) of 0.782 and a counting mean absolute error (MAE) of 3.73, outperforming the original YOLOv11 model by 3.3% in mAP and reducing the MAE from 3.95. The results highlight the model’s effectiveness in high-precision detection and counting of wildlife in complex natural environments, offering a scalable and real-time solution for conservation and ecological monitoring applications.