基于超像素分割的实时野外场景理解

Real-Time Nature Scene Understanding Based on Superpixel Segmentation

提出了一个实时野外场景理解算法,可以广泛地应用于自主野外环境探索、辅助驾驶系统等方面.为了提升分类算法的速度和精度,采用快速的超像素分割来对场景进行预处理,然后针对每个超像素区域提取HSV颜色特征、LBP纹理特征和EOH边缘梯度特征构建其多维特征向量,并采用多类Real AdaBoost算法进行特征训练得到场景分类器.实验证明,提出的算法不仅具有良好的实时性,同时由于采用超像素进行分割预处理,有效地提升了对场景中不同类别的分类精度.

A real-time algorithm for nature scene understanding is proposed, which is able to be used for environment exploration and driving assistance. To improve accuracy and speed, preprocessing is implemented by rapid superpixel segmentation. Combining HSV feature, LBP texture feature and EOH feature to construct feature description, and classifier can be acquired by multi-class Real AdaBoost training. Experimental results prove that the algorithm can meet real time running. Further, classification accuracy is effectively boosted by superpixel segmentation.