面向无人艇环境感知的改进型 SSD 目标检测方法

为了提升无人艇对典型水面小目标感知能力,本文提出了基于多尺度卷积融合结构和空间注意力加强的改进型SSD目标检测算法。首先,对SSD浅层网络进行多尺度卷积融合,提升浅层网络的语义信息;其次,设计空间注意力结构对卷积特征层逐个增强,提升对弱纹理小目标特征保持性;最后,在VOC公开数据集和自构水面目标数据集上进行了测试,并基于无人艇开展了真实海域目标检测识别验证。实验结果表明,该算法在无人艇Nvidia平台的运行效率可达15 fps,能准确检测识别浮标、桥墩、渔船、快艇和货船等目标,在典型海面场景虚警率为5%时的小目标检测率相对原生SSD算法提升近20.2%,平均有效检测率达到79.3%。

Object detection for environment perception of unmanned surface vehicles based on the improved SSD

To improve the perception ability of typical small water targets for unmanned surface vehicle (USV), this paper proposes an improved SSD object detection algorithm based on multi-scale convolution layer fusion and spatial attention enhancement architecture. Firstly, a multi-scale fusion method is utilized to improve the semantic representation of SSD shallow layer for small targets. Secondly, the spatial attention architecture is designed for each convolutional feature extraction layer to improve feature retention of small targets with weak texture. Finally, the proposed algorithm is evaluated on VOC and self-constructed surface target dataset. The real sea target detection and identification verification based on USV are carried out. Experimental results show that the proposed method can reach high operating efficiency with 15 fps on the USV Nvidia platform. The targets, such as buoys, bridge piers, fishing boats, speed boats and cargo ships, can be identified accurately. Compared with the original SSD algorithm, the proposed method could achieve a higher detection rate for small targets in the typical sea scene, which is increased by nearly 20. 2% when the false alarm rate is 5% . The average effective detection rate can reach 79. 3% .