基于视觉采用词袋模型的移动机器人地形分类算法设计  

Visual terrain classification for mobile robot using bag of words

作　　者：孙玉超[1,2] 吴航[1] 苏卫华[1] 陈卓[1] 安慰宁 秦晓丽[1] 

机构地区：[1]军事医学科学院卫生装备研究所,天津300161 [2]天津理工大学中环信息学院,天津300380

出　　处：《医疗卫生装备》2017年第2期114-117,121,共5页Chinese Medical Equipment Journal

基　　金：原总后卫生部重点项目(BWS14C054);军队后勤科研计划重大项目(AWS16J001);天津科技支撑计划重点研发计划(16YFZCSF00590)

摘　　要：目的 :针对野外场景存在多种地形的情况,设计视觉地形分类算法,正确感知所处地形,辅助移动机器人制定合适的运动策略。方法:使用词袋模型和支持向量机构建地形分类算法。词袋模型包括特征提取、码本生成和码本编码。词袋模型输出地形图像的中层特征,特征输入到支持向量机中得到地形分类结果。结果:使用四足机器人平台分别在瓷砖、沥青、沙地和草地环境中进行视觉地形分类实验,实验结果良好,分类平均准确率保持在90%以上;对草地识别性能最好,达到97.54%。结论:该方法能有效准确区分各类地形,准确率高、稳定性好,是一种简单有效的地形分类算法。Objective To design a visual terrain classification algorithm to facilitate the robot to make appropriate movement strategy by perceiving the surrounding environment. Methods Bag of words（BOW） and support vector machine（SVM） were used to develop a simple and effective terrain classification algorithm. The BOW model involved in feature extraction, codebook generation and feature coding. The mid-level feature developed by BOW model was then fed into SVM classifier to obtain the terrain classification result. Results The quadruped robot platform was applied to performing visual terrain classification experiment in the natural environment. The test environment included floor, asphalt, sand and grass. Good experimental results were achieved, and the classification accuracy was above 90%（the beat was 97.54% for grass）. Conclusion The algorithm can effectively and accurately distinguish all kinds of terrains, with high accuracy and good stability. The key frame selection method needs researching in the future.

关 键 词：地形分类 视觉方法 词袋模型 移动机器人 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP242[自动化与计算机技术—计算机科学与技术]