非结构化道路区域检测的协同学习方法

Unstructured road detection using co-learning

非结构化道路区域检测是智能车环境感知的重要问题。提出基于多方向Gabor纹理直方图的SVM分类器,并将其与直方图反向投影器组合,建立了协同学习框架。在实际运行中,两个学习器可以相互为对方提供标注样本进行更新,既提高了在线学习能力,又回避了自学习过程经常导致的模型漂移问题。经实验测试,协同学习机制显著改善了道路检测性能。

Unstructured road region detection forms a main problem of environment sensing for intelligent vehicle. In this paper, two learers are proposed to solve this problem. One is a support vector machine （SVM） classifier which utilizes multi-orientation Gabor texture histogram, and the other is a color histogram back-projection model. Both learners are combined in a co-learning framework. In practical running, the two learners can provide ＂labeled＂ samples for each other. This approach can improve the online learning capability and avoid the model drifting problem which often occurs in self- learning approach. Experimental results show the advantages of the proposed co-learning approach.