Edge-AI Detection and Adaptive Non-Lethal Deterrence of Rhesus Macaques for Nepalese Agriculture Using YOLO26n

Abstract

Rhesus macaque (Macaca mulatta) crop raiding is a
widespread and economically damaging form of human-wildlife
conflict in Nepal, with field studies reporting average crop
losses of 24.62% and raiding frequency peaking at 45% in
spring and 39% during the monsoon. Traditional mitigation
manual guarding, scarecrows, and static acoustic devices
provides no automated early warning and is prone to rapid
habituation, leaving smallholder farmers with limited recourse.
This paper presents the design, implementation, and bench-
level evaluation of an edge-based AI-IoT system that detects
macaques in real time and responds with an adaptive, non-
lethal audio deterrent while logging and notifying farmers of
every event. A YOLO26n object detector, transfer-learned from
COCO-pretrained weights, was trained on a curated single-
class dataset of 10,886 images (8,229 macaque images and 2,657
negative samples of humans, cattle, scarecrows, and empty
fields, added after bench testing exposed human false positives)
and exported to ONNX for CPU inference on a Raspberry Pi 4.
A confidence-threshold decision module (≥ 0.60) triggers a
Bluetooth-driven PAM8403/TS-250 audio actuator with
rotating clips designed to resist habituation, alongside Telegram
and Flutter notifications and an SQLite event log. The
retrained model achieved 89.84% precision, 80.90% recall,
89.9% mAP@50, and 65.75% mAP@50–95, with an average
edge inference latency of approximately 320 ms. Indoor bench
testing confirmed correct end-to-end operation across
detection, deterrence, notification, and logging; field
deployment and quantitative field validation remain the
principal direction for future work.