基于多尺度特征融合的无人机目标检测
首发时间:2020-05-26
摘要:无人机飞行带来的安全问题日益受到关注,在例如机场、军事基地、监狱等许多场景都越来越强调无人机检测的重要性,基于视觉技术的无人机检测能在一定程度上促进无人机反制技术的发展。无人机检测所面临的难点问题包括目标成像尺度变化非常大,飞行背景复杂,通常伴有干扰目标出现。卷积神经网络能够学习到表示能力非常强的目标特征,但网络中低层特征语义信息缺乏,高层特征细节信息不足,致使网络不能对大、中、小目标的检测都具有良好的鲁棒性。本文基于此构建了一个无人机检测数据集,利用Res2net提取目标多感受野特征,提出了一种新的混合特征金字塔结构,从细粒度的多尺度特征提取和层级多尺度特征融合两个方面来提升网络性能,实现了一种基于多尺度特征融合的无人机目标检测网络。在本文构建的无人机数据集上进行实验,提出的网络对无人机、鸟类和普通飞机的识别率均能达到95%以上。
For information in English, please click here
Multi-scale Feature Fusion Based Drone Detection
Abstract:With the growing concern regarding the safety issues of drones, the importance of drone detection is emphasized in many places such as airports, military bases, prisons, and so on. Visual based drone detection can promote the development of drone countermeasures technology to a certain extent. One of the challenges in visual-based drone detection is that imaging scale changes greatly, the background is complex, and the interference target usually appears. Convolutional Neural Network can learn the target features with strong representation ability, but the low-level feature semantic information and high-level feature detail information of the network are insufficient, which makes it unable to detect large and medium-sized targets with good robustness. Based on this motivation, this paper constructed a drone detection dataset, uses Res2net to extract multi-sensory field features, proposed a new hybrid feature pyramid structure improving network performance from fine-grained multi-scale feature extraction and hierarchical multi-scale feature fusion, and designed a multi-scale feature fusion algorithm for drone target detection. Based on the drone dataset constructed in this paper, the discrimination of drones, birds, and airplanes is over 95% in our experiments
Keywords: drone detection image processing computer vision multi-scale feature fusion
引用
No.****
动态公开评议
共计0人参与
勘误表
基于多尺度特征融合的无人机目标检测
评论
全部评论0/1000