Abstract:
Human-wildlife conflict (HWC) in Sri Lanka is a significant issue, as elephants, monkeys,
and peacocks often destroy crops and threaten livelihoods. Conventional control methods
such as electric fencing and patrols remain expensive and reactionary. This research
developed a real-time wildlife detection system based on YOLOv11s, a deep learning model
trained on 5,000 hand-selected and curated images obtained from Sri Lankan habitats. Image
augmentation was applied during data preprocessing, while a disambiguation pipeline
incorporating both animal and human input was established to reduce false alarms.
Validation results showed a mean average precision (mAP@0.5) of 92.1%, and species-
specific accuracies of 94.8%, 96.0%, and 85.6% for elephants, peacocks, and monkeys,
respectively. The system achieved real-time inference processing at 9.8 ms per frame and
incorporated dual alert schemes using local audio alarms and Telegram messages. Compared
to YOLOv8s, YOLOv11s demonstrated 20% higher accuracy and faster processing, making
it suitable for resource-limited conservation applications. This research underscores the
potential of deep learning-based monitoring to minimize agricultural losses, enhance rural
safety, and promote human-wildlife coexistence in Sri Lanka.
