融合超像素聚类与SVM分类的非结构化道路识别算法  

An unstructured road recognition algorithm combining superpixel clustering and SVM classification

作　　者：陈涛 宫金良[1] 张彦斐 CHEN Tao;GONG Jinliang;ZHANG Yanfei(School of Mechanical Engineering,Shandong University of Technology,Zibo Shandong 255049,China;School of Agricultural Engineering and Food Science,Shandong University of Technology,Zibo Shandong 255049,China)

机构地区：[1]山东理工大学机械工程学院,山东淄博255049 [2]山东理工大学农业工程与食品科学学院,山东淄博255049

出　　处：《激光杂志》2022年第1期33-38,共6页Laser Journal

基　　金：国家自然科学基金(No.61303006);山东省重点研发计划项目(No.2019GNC106127);淄博市生态无人农场研究院项目(No.2019ZBXC200)。

摘　　要：为解决移动机器人对非结构化道路识别的准确性、实时性问题,提出了改进的超像素聚类与支持向量机融合的监督修正算法。首先对采集的道路图像进行预处理,仅在此道路的显著性区域内进行图像的平滑处理,然后基于改进的线性迭代聚类算法,将图像分割为内部像素较为一致的若干超像素单元,根据灰度差准则进行超像素的合并,以超像素块作为训练样本集,加快了支持向量机(SVM)分类器的训练学习速度,提取合并后的超像素颜色、纹理特征作为分类器训练集,构造SVM道路分类器并在测试集中进行分类识别,最后根据建立的评估函数对分类结果实时修正,保障了分类识别的准确率。实验表明,该算法的分类效果精确度高,且具有良好的实时性和鲁棒性。In order to solve the problem of recognition accuracy and real-time performance of intelligent vehicles on unstructured roads, an improved superpixel clustering and SVM fusion algorithm is proposed. Firstly, the acquired image was preprocessed, the road image is smoothed within the significance region. Then, based on the improved linear iterative clustering algorithm, the image is segmented into several superpixel units with relatively consistent internal pixels, the superpixels are merged according to the gray difference criterion, Taking the superpixel block as the training sample set, the training learning speed of SVM classifier is accelerated, the merged hyperpixel color and texture features are extracted as the classifier training set, the SVM road classifier was constructed and identified in the test set. Finally, the classification results are modified in real time according to the established evaluation function, which ensures the accuracy of classification recognition. The experimental results show that the algorithm has high classification accuracy, good real-time performance and robustness.

关 键 词：非结构化道路 聚类算法 特征提取 超像素 支持向量机 

分 类 号：TN209[电子电信—物理电子学]