Deep learning technique for Swamp deer detection Using Cost-Effective UAVs

This study enhances drone capabilities for wildlife detection, focusing on swamp deer (Rucervus duvaucelii). We used YOLO V3, V5, V7, V8, Object Detection V3, and DETR models.
We prepared a non-GPU Real-time detection using frame sampling technique, making it cost-effective and accessible, suitable for conservation efforts and adaptable to other species monitoring.

Total images - 8210 UAV Aerial image - 6765 Handheld camera - 1445
UAV utilized - DJI Mavic 2 Zoom, DJI Mavic 2 Enterprise, and DJI Mavic Pro
YOLO V3 Train – 6198, Test – 2012
Others (YOLO V5, V7, V8, DETR, Object detection) Train-6198, Test- 687, Validate- 1325

Real-time - Using frame skipping technique, The YOLO V5 model has shown outstanding performance when applied to video with 19 skipped frames at a resolution of 320 pixels and 32 frames per second (fps).

The Swamp deer dataset was annotated manually using Labelmg offline tools (https://pypi.org/project/labelImg/) and Roboflow online platform (https://app.roboflow.com/)

本研究聚焦泽鹿（Rucervus duvaucelii），旨在提升无人机（UAV）野生动物检测能力。我们采用了YOLO V3、V5、V7、V8、目标检测V3（Object Detection V3）以及DETR模型。
我们开发了一种无需图形处理器（GPU）的实时检测方案，采用帧采样技术，兼具成本效益与易用性，适用于野生动物保护工作，且可适配其他物种的监测任务。

图像总规模：8210张，其中无人机（UAV）航拍图像6765张，手持相机拍摄图像1445张。
所用无人机机型包括：DJI Mavic 2 Zoom、DJI Mavic 2 Enterprise及DJI Mavic Pro。

YOLO V3：训练集6198张，测试集2012张；
其余模型（含YOLO V5、V7、V8、DETR及目标检测模型）：训练集6198张，测试集687张，验证集1325张。

实时检测方案采用帧跳过技术，其中YOLO V5模型在分辨率320像素、帧率32帧每秒（fps）、跳过19帧的视频输入下展现出优异性能。

本泽鹿数据集通过离线工具Labelmg（https://pypi.org/project/labelImg/）与在线平台Roboflow（https://app.roboflow.com/）完成手动标注。