Smart Wildlife Detection Framework Using YOLOv8 Segmentation for Conflict Mitigation

Abstract

Human–wildlife interactions have intensified in many rural and forest-bordering communities due to widespread habitat destruction and ongoing deforestation [1,2,3,4,5]. These encounters frequently cause crop losses, property damage, and serious risks to human lives. Conventional surveillance mechanisms often lack the precision and response speed needed to identify wildlife intrusions effectively, resulting in delayed interventions and increased danger for both people and animals. To address this gap, this work introduces an enhanced YOLOv8 segmentation framework integrated with a CSPDarkNet53 backbone. By partitioning feature maps and minimizing redundant computation, CSPDarkNet53 strengthens feature extraction while preserving essential information. The model employs multi-scale convolutional filters (3×3 and 5×5), batch normalization, and Leaky ReLU for stable learning, along with a PANet-based neck that improves multi-level feature aggregation. Experimental analysis demonstrates superior segmentation and detection performance, achieving a bounding box mAP of 0.86 and mask mAP of 0.79—outperforming YOLOv8-M and YOLOv10-M. High recall values for elephants (0.841) and tigers (0.891) further validate the model's robustness. While segmentation for elephants is consistently accurate, occasional tiger–background misclassifications highlight areas for improvement. Overall, the proposed model offers a practical and efficient approach to wildlife monitoring and conflict mitigation.

Keywords