基于随机蕨丛的3D-SLAM重定位方法

A 3D-SLAM relocalization method based on randomized ferns

针对3D-SLAM系统在未知环境中建图可能会出现的匹配失败而导致定位丢失的问题,提出了一种基于随机蕨丛的视觉里程计实时重定位方法。利用视觉里程计算法获取关键帧和关键帧对应的位置姿态;利用随机蕨丛算法对关键帧图像编码,并存储编码结果;通过定义随机蕨丛编码结果之间的差异,增量存储关键帧;当视觉里程计模块判断当前定位丢失时,比对当前帧与之前关键帧的编码结果,寻找最相近的关键帧进行重定位。采用TUM数据集对算法进行测试,在时间消耗仅增加20%的情况下,实现在3D-SLAM过程中定位丢失时的重定位,提升建图成功率。

To deal with the match failure and further relocation failure during the 3D-SLAM system working under unknown environments, a real-time relocation method is proposed based on visual odometry using randomized ferns. Firstly,the method obtains the keyframes and their corresponding poses by visual odometry. The keyframes are encoded by randomized ferns. The encoding results and the keyframes are saved. Then the keyframes are filtered through defining the differences among the encoding results of each keyframes. The qualified keyframes are used in the randomized ferns training. Finally,when the visual odometry detects the failure of localization,the encoding result of currency frame and the qualified keyframes filtered before are compared to one another to be relocalized by finding the best match. Based on the TUM dataset, the experiment result shows that the 3D-SLAM system can successfully relocate when localization fails. It significantly raises the mapping success rate with only20% time consumption increased.