低成本自运行鸟类监测系统，助力家庭公民科学观测鸟类多样性
Autonomous AI Bird Feeder for Backyard Biodiversity Monitoring
- 用触发式摄像头+本地服务器实现无云隐私保护的自动采集
- 针对40种比利时鸟类的分类准确率达88%（顶1）
- 小入口设计防鸽子，检测裁剪提升分类性能
本文提出一种低成本、本地部署的自主鸟类监测系统，用于比利时城市庭院的鸟类多样性观测。运动触发的IP摄像头通过FTP将短视频上传至本地服务器，帧被采样后使用Detectron2定位鸟类；裁剪区域由在40种比利时鸟类子集上微调的EfficientNet-B3模型进行分类，该子集源自更大的Kaggle数据集。所有处理均在无独立显卡的通用硬件上完成，保障隐私并避免云端费用。物理喂食器采用30毫米小入口，排除鸽子并减少误触发。基于检测引导的裁剪显著提升分类准确率。分类器在筛选后的子集上验证准确率约99.5%，在未见物种上实际应用准确率（顶1）约为88%，证明了家庭级公民科学生物多样性记录的可行性。
原文摘要 · Abstract (English)
This paper presents a low cost, on premise system for autonomous backyard bird monitoring in Belgian urban gardens. A motion triggered IP camera uploads short clips via FTP to a local server, where frames are sampled and birds are localized with Detectron2; cropped regions are then classified by an EfficientNet-B3 model fine tuned on a 40-species Belgian subset derived from a larger Kaggle corpus. All processing runs on commodity hardware without a discrete GPU, preserving privacy and avoiding cloud fees. The physical feeder uses small entry ports (30 mm) to exclude pigeons and reduce nuisance triggers. Detector-guided cropping improves classification accuracy over raw-frame classification. The classifier attains high validation performance on the curated subset (about 99.5 percent) and delivers practical field accuracy (top-1 about 88 percent) on held-out species, demonstrating feasibility for citizen-science-grade biodiversity logging at home.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。