基于FAST角点和仿射改进的随机蕨丛的单目视觉实时匹配算法

A Monocular-vision Real-time Matching Algorithm Based on FAST Corners and Affine-improved Random Ferns

针对传统配准方法在自然特征粗匹配过程中在线匹配速度较慢、识别率较低的问题,提出一种基于FAST角点与仿射改进的适用于单目视觉实时定位的随机蕨丛匹配方案.该方案基于随机蕨半朴素贝叶斯非层次结构分类模型,突破了传统匹配方案中离线过程与在线过程的对称处理架构.采用FAST角点作为环境自然特征提升检测速度,进而改进随机蕨丛离线过程中稳定点选取、训练片元生成的仿射策略,提升识别率、缩短离线训练时间,最后收缩随机蕨丛规模,缩短在线匹配时耗.室内、外实时匹配定位实验结果表明该方案满足单目视觉定位的实时性、识别率需求.

For classic calibration methods, there exist problems of low online matching speed and recognition rate in rough matching of natural features. To solve the problem, a random ferns matching algorithm based on FAST （Features from Accelerated Segment Test） corners and affine-improvement is proposed to realize monocular-vision real-time localization. This algorithm is based on a random fern semi-Bayes nonhierarchic classification model, which breaks the symmetrical framework of on-line and off-line processes in classic matching methods. It adopts FAST corners as the environment natural features to accelerate online detection, improves affine strategy for stable point set selection and training fragment generation in random ferns off-line process to improve recognition rate and reduce off-line training time, and scales back the size of random ferns to reduce time consumption of online matching. Indoor and outdoor matching and localization experiments show that the proposed algorithm meets requirements of real-timeness and recognition rate for monocular-vision localization.