盲人视觉辅助应用中的自然场景文字有效提取仍然是一个亟待解决的问题。自然场景中文字呈现的多样性、盲人在行走过程中的头部旋转以及非正面拍摄引起的文字几何失真等因素，都构成了这一应用中文字提取技术的瓶颈。本项目拟研究：（1）基于BOVW模型的文字标牌检测的高效学习方法；（2）检测-跟踪-模版更新相结合的文字标牌跟踪方法和穿戴式视觉对跟踪的稳定性影响；（3）基于不变特征约束的标牌文字透视矫正估计方法。本课题将理论分析和实验研究相结合，旨在建立自适应决定自然场景目标种类个数的非监督字典学习模型，探索体现文字标牌特点的强区分性特征，实现盲人行走过程中文字标牌的实时稳定跟踪，消除非正面拍摄造成的文字几何失真，为盲人的视觉辅助应用奠定理论基础。

Efficient extraction of natural scene text is still a problem to be solved in the blind visual aid application. The diversity presented by natural scene text, the head rotation of the blind during walking and the geometric distortion resulted by not frontal shooting of camera, have become the bottleneck factors of text extraction technology in the wearable vision application. The research work mainly include: 1) An efficient learning method in the text sign detection is studied, which is based on BOVM model; 2) A TUD (Tracking-Updating-Detection) text sign tracking method is proposed, and the influence of wearable vision on the stability of tracking is studied. 3) The perspective correction of text is estimated under the invariant feature constraint. By means of the theoretical analysis and experiments studies, our objectives are as follows. An unsupervised dictionary learning model is constructed, which aims to adaptively determine the number of scene target species. The strong distinguishing feature is selected by the detector, which can reflect the characteristics of text signs. The real time stable tracking of text sign in the course of the blind walking is realized. The geometry distortion of the text in the walking process of the blind is eliminated. The above research achievement will provide the theoretical basis for the visual aid application of the blind.