YOLO-FES: An Improved Elephant Intrusion Detector Based on YOLOv8n

Abstract

Human–elephant conflict (HEC) poses a significant threat to both biodiversity and rural livelihoods, necessitating innovative monitoring solutions that are both accurate and deployable in resource-constrained environments. Existing deep learning models often trade off detection accuracy for computational efficiency, limiting their real-time use on edge devices. To address this challenge, we propose YOLO-FES, a lightweight object detection model tailored for efficient and accurate elephant intrusion detection. YOLO-FES integrates three key components: (i) a FasterNet block (F) that replaces bottleneck blocks in the backbone, (ii) an Efficient Multi-scale Attention (E) module to enhance feature representation, and (iii) a Slim-neck (S) powered by GSConv for improved feature fusion. The model was trained on a diverse dataset comprising trap cameras, CCTV, drones, mobile phones, television footage, and supplemented with elephant images from COCO2017. Experimental results demonstrate that YOLO-FES reduces parameters, FLOPs, and model size by 19%, 23.5%, and 16.1%, respectively, compared to YOLOv8n, while achieving higher accuracy with +1.5% mAP@0.5 and +1.2% mAP@0.5–0.95. Edge deployment evaluations confirm real-time performance, with inference times ranging from 24.7 ms on Jetson Orin Nano to 254.2 ms on Raspberry Pi 4B. These results establish YOLO-FES as a robust, low-cost, and deployable solution for real-time elephant intrusion detection, contributing to sustainable mitigation of human–elephant conflict.