基于棒状像素与随机森林的道路场景理解

Road Scene Understanding Using Stixel and Random Forest Classifier

本文针对棒状像素(Stixel)在表达障碍物时缺少语义信息的问题,提出了一种生成棒状像素级类别标签的方法。通过颜色、纹理和几何高度特征的联合,推断出像素级语义标签,即障碍物大类、植物与天空。将棒状像素位置信息融合至像素级语义图,用类别标签来优化每个棒状像素的顶点估计后,后以其内部像素的优势类来判定类别,得到语义棒状像素图。分类器选用能够快速推理且具抗过拟合能力的随机森林(RF)分类器。实验结果表明,本文提出的方法能够有效地对棒状像素进行语义表达,在一定程度上优化棒状像素在纹理缺失区域的高度估计。

In order to solve the problem of lacking semantic information when Stixel is used to express obstacles,this paper proposes a method to generate category labels of rod-like pixel level.Through the combination of color,texture and geometric height features,the pixel-level semantic labels,namely obstacle categories,plants and the sky,are inferred.The position information of the bar pixel is fused to the semantic bar pixel graph.After the vertex estimation of each bar pixel is optimized by the category label,the category is determined by the advantage class of its internal pixels,and the semantic bar pixel graph is obtained.The random forest(RF)classifier with fast reasoning and resistance to overfitting was selected.Experimental results show that the method proposed in this paper can effectively express the semantic meaning of rod-like pixels and optimize the height estimation of rod-like pixels in the texture missing region to some extent.