无人机
人工智能
计算机科学
计算机视觉
传感器融合
融合
遥感
特征提取
钥匙(锁)
目标检测
领域(数学)
模式识别(心理学)
图像融合
匹配(统计)
跟踪(教育)
出处
期刊:
日期:2026-08-01
卷期号:: 599-602
标识
DOI:10.1109/icaita71179.2026.11737055
摘要
Monitoring non-cooperative drones in crowded 2.4/5.8 GHz ISM bands is critical for low-altitude security. High-power Wi-Fi/Bluetooth interference makes classification difficult, especially under sensor dropouts. Standard classifiers fail when receiver configurations change. We resolve this with a dual-branch framework combining a residual spectrogram CNN and a temporal MLP branch, fused via a static configuration-dependent routing lookup table. For semi-supervised learning on unlabeled data, we use a 0.99 confidence threshold for the CNN to reject noise and a rank-based top-800 selection for the MLP to prevent configuration starvation. On the World AI+Radio Challenge 2026 dataset, this design achieves 99.21% transductive test accuracy and 99.18% ± 0.08% validation accuracy, significantly outperforming the 98.91% inductive baseline. The minimal validation-to-test gap confirms representation robustness without leakage, demonstrating stable classification during sensor outages.
科研通智能强力驱动
Strongly Powered by AbleSci AI