视觉导盲仪是一种旨在解决盲人出行困难的导盲设备, 为了实现视觉导盲仪诱导盲人找到盲道并沿盲道行走, 提出了一种基于机器学习识别与标记分水岭分割的盲道图像定位算法, 通过离线训练与在线的识别、分割来定位盲道区域。首先对盲道图像进行视角变换的预处理, 根据盲道的地面方程将变化的倾斜视角转换为固定的俯视视角, 消除射影变换带来的失真; 然后利用局部二进制模式描述子提取鸟瞰图的纹理特征, 以自适应增强学习算法离线训练盲道识别分类器; 进而利用分类器对鸟瞰图像进行在线识别, 粗略确定盲道区域; 最后将识别结果进行形态学处理后作为标记, 利用标记分水岭算法得到精确分割的盲道区域并定位盲道中心线。在研制的视觉导盲仪上进行验证, 结果表明盲道定位准确率达到了95.44%, 速度平均每秒8帧, 具有高准确率的同时达到实时性要求, 为盲道的准确三维定位提供了必要条件。

VTA (Visual Travel Aids) are devices used for addressing traveling difficulties of visually impaired people. To develop VTA for guiding visually impaired people to blind sidewalks, a method for blind sidewalk image location was presented based on machine learning recognition and a marked watershed algorithm. The algorithm located blind sidewalks by combining offline training with online recognition and segmentation. First, a blind sidewalk image was pretreated by converting an original image from a camera gradient view into an aerial view based on the plane equation of the blind sidewalk. The pretreating eliminates distortions. A Local Binary Pattern (LBP) descriptor then extracted the texture features of the aerial-view images. An offline cascade classifier trained through Adaboost recognized the blind sidewalk based on the LBP descriptor. The cascade classifier recognized the aerial view image online and roughly identified the blind area. The recognition results were then morphologically processed as markers to obtain the exact segmentation of the blind area through a marked watershed algorithm. Finally, the segmentation results were used to locate the centerline of the blind sidewalk. The algorithm was validated on the VTA. The experimental result showed that the blind sidewalk localization achieved 95.44% accuracy with an average speed of 8 frame/s. It exhibited a high accuracy rate while satisfying the real-time requirement, which are the necessary conditions for accurate 3D localization of blind sidewalks.