计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (16): 204-212.DOI: 10.3778/j.issn.1002-8331.2201-0067

• 图形图像处理 • 上一篇    下一篇

基于优化SSD的低空无人机检测方法

张灵灵,王鹏,李晓艳,吕志刚,邸若海   

  1. 1.西安工业大学 兵器科学与技术学院,西安 710021
    2.西安工业大学 发展规划处,西安 710021
    3.西安工业大学 电子信息工程学院,西安 710021
  • 出版日期:2022-08-15 发布日期:2022-08-15

Low-Altitude UAV Detection Method Based on Optimized SSD

ZHANG Lingling, WANG Peng, LI Xiaoyan, LYU Zhigang, DI Ruohai   

  1. 1.School of Ordnance Science and Technology, Xi’an Technological University, Xi’an 710021, China
    2.Development Planning Service, Xi’an Technological University, Xi’an 710021, China
    3.School of Electronic Information Engineering, Xi’an Technological University, Xi’an 710021, China
  • Online:2022-08-15 Published:2022-08-15

摘要: 针对浅层特征缺乏语义信息和小目标特征不显著的问题,提出了一种基于多尺度特征融合和注意力的低空无人机(unmanned aerial vehicle,UAV)检测方法。首先提出一种多尺度特征融合模块,将不同尺度的特征图进行有效融合,使浅层特征图的细节纹理信息和深层特征图的语义信息得到充分的利用,改善浅层特征语义信息不足的问题。然后在网络特征图输出处引入一种不降维局部跨信道交互策略和核大小自适应选择的通道注意力机制,以极其轻量级的方式获取跨通道的交互信息。为使先验框和有效感受野匹配,优化默认框设置方法,更好地检测小目标。使用自制无人机数据集进行验证,结果表明改进后算法平均准确率为84.07%,比原始SSD(single shot multibox detector)算法提高了7.81个百分点,检测速度达到31.3?frame/s。

关键词: 目标检测, 多尺度特征融合, 低空无人机, 注意力机制

Abstract: For the problems of shallow features lacking semantic information and small target features not significant, a low-altitude UAV(unmanned aerial vehicle) detection method based on attention and multi-scale feature fusion is proposed. Firstly, a multi-scale feature fusion module is proposed to effectively fuse the feature maps of different scales. The detailed texture information of shallow features map and semantic information of deep features map are fully utilized to improve the lack of semantic information of shallow features. Then a local cross-channel interaction strategy without dimensionality reduction and a channel attention mechanism for adaptive selection of core size are introduced at the output of the network feature graph to obtain cross-channel interaction information in an extremely lightweight way. Finally, in order to match the prior boxes with effective receptive field, the default box setting method is optimized to better detect small targets. The self-made UAV data set is used for verification. The results show that the average accuracy of this method is 84.07%, which is 7.81 percentage points higher than original SSD(single shot multibox detector) algorithm, and the detection speed is 31.3 frame/s.

Key words: object detection, multi-scale feature fusion, low-altitude UAV, attentional mechanism