一种双目混合追踪的视觉里程计算法

Hybrid Visual Odometry Algorithm For Stereo Camera

针对特征点法视觉里程计(visual odometry,VO)中,描述子匹配算力要求较高,而光流追踪算法存在误差累积的问题,提出了一种双目混合追踪的视觉里程计算法,并通过实验对该方法的定位结果与计算效率进行分析。在该算法中,前端在影像关键帧提取ORB(oriented brief)特征点并生成路标点,优先采用Inverse Compositional光流进行追踪,使用卡方检验等检测外点;为检测定位粗差,提出通过由相邻帧位姿恢复的极线进行检验;光流追踪失败或未通过极线检验时,提取ORB特征点维持跟踪或进行重定位;同时,后端进行相机位姿、路标点联合优化以减小误差累积。实验结果表明,该算法定位精度低于ORB-SLAM2(simultaneous location and mapping)算法,但运算效率有明显提升。

Considering the heavy computation requirment of descriptors matching in key point visual odometry,and the accumulated error caused by optical flow tracking,we propose the hybrid tracking algorithm for stereo camera and analyze its orientation results and calculation efficiency through experiments.In the front end,ORB feature points are extracted in the key frames of images and generated as waypoints and tracked via inverse compositional optical flow,then chi-square test is used to check exterior points.To check positioning gross error,the restored polar curve of relevant adjacent frame’s pose is suggested to use.Meanwhile,in the back end,the incremental nonlinear optimization such as camera’s pose has been done to reduce error accumulation.Experiments prove that the hybrid algorithm has lower localization accuracy but higher calculation efficiency compared to the current open source ORB-SLAM2 algorithm.