Answers are grounded in this paper and cite their sources.
Rescue teams operating beneath a forest canopy may be only partly visible in aerial images. Recognising their actions is therefore difficult precisely where visual monitoring could assist coordination during forest fires and post-disaster searches. Foliage hides body parts, illumination varies across the scene, and background textures can resemble the remaining target features. We develop OFRA-YOLO (Occlusion-aware Forest Rescue Action YOLO) around YOLOv8n for this setting, where YOLO stands for You Only Look Once. Its feature-fusion stage uses a weighted bidirectional feature pyramid network (BiFPN), so the contribution of each scale is learned during training. Large Separable Kernel Attention (LSKA) is placed on the P5 detection branch to incorporate spatial context beyond the immediately visible body region. For box optimisation, Wise Intersection over Union (WIoU) replaces Complete Intersection over Union (CIoU) and adjusts sample contributions through non-monotonic focusing. The three architectural and loss-function changes address scale variation, background interference, and uneven localisation difficulty, respectively; controlled Cutout is used during training. On a forest rescue-action dataset containing 8601 images, the final OFRA-YOLO configuration achieves 80.78% mean Average Precision (mAP@0.5), 74.87% precision, and 78.06% recall. The mAP@0.5 score exceeds that of YOLOv8n by 9.73 percentage points. The results support the use of this compact detector for recognising rescue actions in cluttered forest imagery.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: