中文摘要：

基于视觉的智能车辆导航技术是通过对各种道路环境进行感知和理解，从而确定智能车辆的可行驶区域。针对实际道路环境的复杂性与多样性问题，提出了能够适应复杂环境的道路识别算法。首先，使用SLIC（simplelineariterativeclustering）超像素分割算法对原始道路图像进行超像素分割，得到性质相同、尺寸均匀的超像素块；其次，基于超像素块使用K—means聚类算法提取出图像中道路区域与非道路区域的K维特征数据，并将提取的特征数据组成训练数据集；然后，针对经典双支持向量机（TSVM）训练时间久、无法求解逆矩阵的问题进行适当矫正，使用训练数据集训练矫正后的双支持向量机；最后，使用训练好的双支持向量机进行道路与非道路的分类识别。四组道路场景的实验结果表明，与基于滑动窗口和颜色、Gabor纹理特征的方法进行对比，该算法能够有效地识别具有阴影、水迹、障碍物等复杂环境下的道路；以人工标注结果为标准，前三组识别错误率低于0．1，第四组识别错误率低于0．15；与传统SVM相比，矫正的TSVM具有更高的效率，可以大大降低训练时间。该算法在复杂环境下道路识别错误率低，性能良好，为道路环境感知和理解提供了一种新的方法。

英文摘要：

Intelligent vehicle navigation technology based on computer vision can aware and understand road environment, so it can help intelligent vehicle to find a valid path. In view of the complexity and diversity of actual road environment, the tradi- tional method often can＇ t cope with the interference factors of actual road,such as,shadow,water stains and leaves. This paper proposed a new algorithm. Under the condition of shadow, water stains and leaves, it could distinguish between road and not road in the different road images. Firstly, using SLIC algorithm to segment original road images, divided the original road images into irregular image patches in which the pixels were homogenous and at the same illumination level. The SLIC algorithm pro- duced high quality segmentations. Secondly, it extracted K-dimension features of road super pixels and not road super pixels by the means of K-means clustering. Then,the training set was made up of K-dimension features. Thirdly,in consideration of the weakness of the TSVM （ twin support vector machine ） , for example, long training time and unable to solve inverse matrix, it added the regularization item and changed the inequality constraints to equality constraints. It used improved TSVM to run training in the training set. Finally, the improved TSVM classified the road and not road in the road images. Compared with the methods based on sliding window and color, Gabor texture features, four group of the experimental results show that it can effec- tively identify the road and not road in the road image with shadow, water stains and leaves. In addition to, the algorithm＇ s per- formance is compared with manually annotated frames to measure the accuracy. The error rate of three groups is lower than 0. 1 ,the error rate of one group is lower than 0.15 ,and improved TSVM has higher execution efficiency. For the road recogni- tion problem under complex environment,it has a good performance and provide a new method for intelligent vehicle navigation based on computer vis