Robust Wildlife Detection Using Objectness-Aware Yolo with Attention and Pose Estimation

In recent years, object detectors based on deep learning have significantly advanced wildlife monitoring by automating identification of animals in diverse and challenging environments. However, existing models, including You Only Look Once (YOLOv3) and Objectness Aware YOLO (OAYOLO), struggle to detect small, occluded, or visually complex animals because of limited spatial focus, fixed label assignment, and lack of structural awareness. To address these limitations, an Attention and Pose-Aware OA-YOLO (APA-OA-YOLO) framework is proposed for robust wildlife detection. Initially, the model used a custom-built wildlife dataset, followed by enhanced spatial discrimination by integrating Convolutional Block Attention Modules (CBAM) into the DarkNet53 backbone and a feature fusion neck. Furthermore, the objectness-aware decoupled head is retained to introduce a refined objectness definition combining the Intersection over Union (IoU) and classification confidence. Subsequently, the natural break label assignment (NBLA) strategy is applied for adaptive anchor classification, incorporating assistant samples into objectness-weighted classification loss. Furthermore, a lightweight pose estimation branch predicts animal key points and uses pose information with detection features for refined bounding box prediction. The experimental results demonstrate that APA-OA-YOLO achieves an improved detection performance with an Average Precision (AP) of 67.8 %, True Positive Rate (TPR) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 4. 1 \%}$</tex>, and recall of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 5. 1 \%}$</tex>.