SIA OpenIR  > 光电信息技术研究室
面向无人机的地面车辆识别算法研究
Alternative TitleResearch on Vehicle Recogonation Algorithm Towards Unmanned Aerial Vehicle
张钟毓
Department光电信息技术研究室
Thesis Advisor刘云鹏
Keyword无人机 卷积神经网络 目标识别 注意力机制
Pages67页
Degree Discipline模式识别与智能系统
Degree Name硕士
2020-05-26
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract使用无人机进行图像采集与目标识别在军事领域得到了越来越广泛的关注,但是由于无人机的拍摄高度距离目标较远、拍摄视角均为俯视等原因,造成了图像中地面目标的体积较小、形状扁平、树木或房屋遮挡情况较严重,现有的目标识别算法应用在上述场景中效果较差。本文着重对面向无人机的地面车辆目标的快速、准确识别进行研究,论文的主要工作成果如下:首先,通过分析面向无人机的目标识别任务的发展历程与任务特点,总结出本课题的重点关注方向,为进一步开展面向无人机的地面车辆目标识别算法研究提供了依据;其次,现有的目标识别算法大多使用的是人工设计的特征,在面向无人机的目标识别任务时准确度较差、速度较慢。针对以上情况,我们设计了一种新的基于深度学习的单阶段目标识别网络模型DRFP。该模型以残差结构为骨架,使用特征金字塔结构实现特征融合;同时在交叉熵损失函数中添加了调整因子,从而实现了对难样本的重点关注;并且使用高斯型非极大值抑制算法提高目标密集区检出率。在无人机航拍数据集上的综合评价结果表明,该模型具有较好的识别速度,但是精度没有达到两阶段模型的水平。最后,我们在上述模型的基础上,将注意力机制与特征融合相结合,大幅提高了识别准确率,但同时造成了识别速度的下降。为了实现实时性识别的目标,我们使用模型压缩的方法进行改进,并最终提出了一种无人机航空影像车辆目标实时识别模型DAGN。在多个无人机航拍图像数据集上的评估结果表明,该方法具有较好的识别准确度和速度。
Other AbstractThe use of UAV (Unmanned Aerial Vehicle) for image acquisition and target recognition has become more and more widely used in the military and civil fields. However, due to the high shooting height and the overhead view of UAV, the ground target in the image are volume smaller, shapes flatter, and they are always obscured by trees or houses. The current target recognition algorithms are not effective in the above scenarios. This thesis focuses on the fast and accurate recognition of ground vehicle targets collected by UAV. The main contents of the thesis are as follows. Firstly, by analyzing the development history and the characteristics of the target recognition task towards UAV, the main focus of this topic is summarized, which provides a basis for further research on the vehicle target recognition algorithm for the UAV. Secondly, most of the existing target recognition algorithms use artificially designed features, which are poor accuracy and slow in target recognition tasks for UAV. In view of the above situation, we have designed a new deep learning single-stage target recognition network named DRFP. This model uses the residual structure as the skeleton, and uses the feature pyramid structure to achieve feature fusion. At the same time, the cross-entropy function with adjustment factors is added to the loss function to achieve the focus on difficult samples, and uses Gaussian non-maximum suppression algorithm to improve the detection rate of the dense area. The comprehensive evaluation results on the VEDAI (Vehicle Detection in Aerial Imagery) data set show that the model has better recognition accuracy. Finally, in response to the requirements of real-time target recognition, we use model compression based on the DRFP model. But at the same time, model compression causes a decrease in recognition accuracy. After combining the attention mechanism and feature fusion, the recognition accuracy was greatly improved, and a real-time vehicle target recognition model for UAV aerial image was finally proposed. Evaluation results on multiple UAV aerial image data sets show that the method has better recognition accuracy and recognition speed.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27119
Collection光电信息技术研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
张钟毓. 面向无人机的地面车辆识别算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
面向无人机的地面车辆识别算法研究.pdf(3053KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张钟毓]'s Articles
Baidu academic
Similar articles in Baidu academic
[张钟毓]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张钟毓]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.