引用本文:张琪安,张波涛,吕强,王亚东.采用卷积神经网络的低风险可行地貌分类方法[J].控制理论与应用,2020,37(9):1944~1950.[点击复制]
ZHANG Qi-an,ZHANG Bo-tao,Lu Qiang,Wang Ya-dong.Low-risk terrain classification based on convolutional neural network[J].Control Theory and Technology,2020,37(9):1944~1950.[点击复制]
采用卷积神经网络的低风险可行地貌分类方法
Low-risk terrain classification based on convolutional neural network
摘要点击 1791  全文点击 579  投稿时间：2019-08-19  修订日期：2020-04-24
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2020.90688
  2020,37(9):1944-1950
中文关键词  移动机器人  地貌识别  低风险地貌  卷积神经网络
英文关键词  mobile robot  terrain classification  low-risk terrain  convolutional neural network
基金项目  
作者单位E-mail
张琪安 杭州电子科技大学 zhang qaya@163.com 
张波涛* 杭州电子科技大学 billow@hdu.edu.cn 
吕强 杭州电子科技大学  
王亚东 杭州电子科技大学  
中文摘要
      针对现有识别方法中风险地貌误判率高、手动地貌特征提取具有局限性等问题, 提出了用于室外移动机器 人的低风险地貌识别策略. 该策略以降低移动机器人遇险率为高优先级目标, 采用双重验证策略, 首先采用多分类 器对所有地貌进行识别, 其后使用二分类器对多分类结果中的安全地貌再次鉴别. 基于该策略, 分别设计了2个卷积 神经网络(CNN), Terrain–CNNⅠ用于多分类识别, Terrain–CNNⅡ则用于二分类安全确认. 为解决地貌样本相对稀缺 问题, 收集了包含水面、草地、泥地、柏油路、沙地、碎石路共6类地貌图像, 通过数据增强方式快速扩充数据集用于 网络的训练与测试. 实验结果表明: 所述方法在维持整体地貌识别率很高的前提下, 显著降低了关键危险地貌的误 判率.
英文摘要
      To deal with the limitations of manual feature extraction and high misjudgement rate of risky terrain in existing recognition methods, a low-risk terrain recognition strategy is proposed to classify the outdoor terrains for mobile robots. This strategy, regarding risk reduction as the primary objective, is dual-verificated. Firstly, a multiple classifier is used to identify all terrains, and then a binary classifier is used to identify the safe terrains in the multi-classification results to reduce the misjudgment of risk terrain. Based on this strategy, two convolutional neural networks are designed, the Terrain–CNNⅠfor multi-classification of recognition, and the Terrain–CNNⅡfor two-classification of safety confirmation. In addition, in order to solve the problem of terrain samples, six kinds of terrain images including water surface, grassland, mud, asphalt, sand and gravel road are collected. After data enhancement, these samples are divided into training samples and testing samples for networks’ training and testing. The experimental results show that the proposed method can significantly reduce the error rate of critical risky terrains while maintaining a high average accuracy.