基于深度特征提取的无人机检测算法

Drone Detection Based on Depth Feature Extraction

随着无人机技术在生产生活中的广泛应用,由其引起的公共安全、隐私保护等问题也日渐突出,因此基于计算机视觉的无人机检测技术逐渐成为当下的研究热点。目前,常用的深度学习目标检测方法如Faster-RCNN、Yolo等在通用的目标检测领域已经可以获得良好的检测性能。但是在无人机检测任务上,由于目标小、边缘设备算力低等限制因素,常用的目标检测算法无法有效地应对这些难题。对此,文中基于轻量级无人机检测网络TIB-Net,引入改进的特征融合模块,将层间特征金字塔模型与像素洗牌方法结合并集成到主干网络,提出了基于深度特征提取的无人机检测算法。该算法不仅通过像素洗牌增强了小目标的细节特征,同时由于层间金字塔的引入,扩大了深度网络的感受野,增强了网络特征提取能力。最终在无人机数据集上进行了测试,对比结果显示该算法对无人机检测的效果有明显提升,最终mAP达到90.4%。

With the widespread application of the drone technology in daily life,the issues about public safety,privacy protection and others caused by it have become increasingly prominent.Therefore,drone detection technology based on computer vision has gradually become a research hotspot.At present,most object detection methods based on deep learning such as Faster-RCNN,Yolo and others have achieved outstanding performance in the common object detection field.However,due to the limitation factors of small target and low computing power of edge devices,the commonly used object detection algorithms cannot effectively deal with these problems.In this regard,based on the lightweight drone detection network TIB-Net,we propose an improved super-resolution inter-layer feature fusion algorithm by combining pixel shuffle method with inter-layer feature pyramid structure and integrating it into backbone.This algorithm not only enhances the detailed features of small targets through the super-resolution method,but also expands the receptive field of the deep network due to the inter-layer pyramid structure,and greatly enhances the feature extraction ability of the network.After testing on the drone dataset provided by TIB-Net,the comparative results show that the effect of the proposed algorithm in UAV detection has been significantly improved,and the final mAP reaches 90.4%.