Journal of Computer Applications ›› 2017, Vol. 37 ›› Issue (7): 2008-2013.DOI: 10.11772/j.issn.1001-9081.2017.07.2008

Previous Articles     Next Articles

Field scene recognition method for low-small-slow unmanned aerial vehicle landing

YE Lihua1,2, WANG Lei1, ZHAO Liping1,3   

  1. 1. College of Electronics and Information Engineering, Tongji University, Shanghai 201804, China;
    2. College of Mathematics Physics and Information Engineering, Jiaxing University, Jiaxing Zhejiang 314000, China;
    3. Institute of Very Large Scale Integration, Tongji University, Shanghai 200092, China
  • Received:2016-12-09 Revised:2017-02-27 Online:2017-07-10 Published:2017-07-18
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61601200).

低小慢无人机降落野外场景识别方法

叶利华1,2, 王磊1, 赵利平1,3   

  1. 1. 同济大学 电子与信息工程学院, 上海 201804;
    2. 嘉兴学院 数理与信息工程学院, 浙江 嘉兴 314000;
    3. 同济大学 超大规模集成电路研究所, 上海 200092
  • 通讯作者: 叶利华
  • 作者简介:叶利华(1978-),男,浙江衢州人,讲师,博士研究生,主要研究方向:计算机视觉、图形图像处理;王磊(1961-),男,陕西西安人,教授,博士,主要研究方向:智能控制、导航与制导;赵利平(1984-),女,湖南衡阳人,博士研究生,CCF会员,主要研究方向:视频编码算法。
  • 基金资助:
    国家自然科学基金资助项目(61601200)。

Abstract: For the complex and autonomous landing scene is difficult to be recognized in wild flight environment for low-small-slow Unmanned Aerial Vehicles (UAV), a novel field scene recognition algorithm based on the combination of local pyramid feature and Convolutional Neural Network (CNN) learning feature was proposed. Firstly, the scene was divided into small scenes of 4×4 and 8×8 blocks. The Histogram of Oriented Gradient (HOG) algorithm was used to extract the scene features of all the blocks. All the features were connected end to end to get the feature vector with the characteristics of spatial pyramid. Secondly, a depth CNN aiming at the classification of scenes was designed. The method of tuning training was adopted to obtain CNN model and extract the characteristics of deep network learning. Finally, the two features were connected to get the final scene feature and the Support Vector Machine (SVM) classifier was used for classification. Compared with other traditional manual feature methods, the proposed algorithm can improve the recognition accuracy by more than 4 percentage points in data sets such as Sports-8, Scene-15, Indoor-67 and a self-built one. The experimental results show that the proposed algorithm can effectively improve the recognition accuracy of the landing scene.

Key words: Convolutional Neural Network (CNN), feature extraction, Unmanned Aerial Vehicle (UAV), scene classification, pyramid model

摘要: 针对低小慢无人机野外飞行场景复杂自主降落场景识别问题,提出了一种融合局部金字塔特征和卷积神经网络学习特征的野外场景识别算法。首先,将场景分为4×4和8×8块的小场景,使用方向梯度直方图(HOG)算法提取所有块的场景特征,所有特征首尾连接得到具有空间金字塔特性的特征向量。其次,设计一个针对场景分类的深度卷积神经网络,采用调优训练方法得到卷积神经网络模型,并提取深度网络学习特征。最后,连接两个特征得到最终场景特征,并使用支持向量机(SVM)分类器进行分类。所提算法在Sports-8、Scene-15、Indoor-67以及自建数据集上较传统手工特征方法的识别准确率提高了4个百分点以上。实验结果表明,所提算法能有效提升降落场景识别准确率。

关键词: 卷积神经网络, 特征提取, 无人机, 场景分类, 金字塔模型

CLC Number: