基于特征交叉检验的实时视觉里程计方法

在自动驾驶和机器人导航系统中，里程计是用于持续获得系统姿态信息的一种装置。视觉里程计能以较低代价获得高精度的目标移动轨迹，基于特征的视觉里程计方法具有时间复杂度较低、计算速度快的优势，有助于数据实时处理。然而，传统基于特征的视觉里程计方法面临着2个技术瓶颈：特征匹配的准确度不足；姿态解算中目标函数的权重值有效性低。为了解决帧间特征匹配准确度不足的问题，本文提出特征交叉检验闭环匹配策略，即在传统单向闭环匹配的基础上，增加反向验证的过程，以获得匹配准确度更高的匹配点集合。该策略解决了传统特征匹配中使用单向闭环匹配策略鲁棒性不足、内点比例低的缺陷，提高了解算精度。同时在交叉检验匹配策略中利用前一时刻的运动信息缩小当前时刻特征匹配的搜索范围，降低特征点匹配的时间复杂度。针对目标函数的权重值有效性低的问题，本文将特征点在图像序列中的出现次数作为其生存周期，提出基于特征点生存周期的目标函数权值设置方法。在姿态解算中，特征点的生存周期可以有效反映其稳定性，使用其作为目标函数权值可以降低解算过程中的累积误差。本文在公开的KITTI数据集中进行算法测试，实验结果证明该方法可以实现高精度、实时的视觉里程计算。

A real-time visual odometry method based on crosscheck of feature

Odometry is widely applied for continuously obtaining system poses in automatic drive system and robot navigation system. Visual odometry can achieve high precision of target motion trajectory estimation with low cost, while feature-based visual odometry has the advantages of low time complexity and high processing speed which are conducive to real-time processing. However, traditional feature-based visual odometry has two technical bottlenecks:low accuracy of feature detection and matching, and the low effectiveness of objective function weight in pose estimation. To address the low accuracy for the feature matching between frames, we present the crosscheck feature matching strategy. It adds the reverse check on the foundation of traditional single-track ‘circle’ matching strategy to obtain more accurate matching feature sets. This strategy increases inlier ratio and solves the low robustness problem in a single-track ‘circle’ strategy, which improves estimation accuracy. Meanwhile, we use motion information of previous frame to reduce the searching scope of current frame in crosscheck strategy. To address the low effectiveness of objective function weight, we use the occurrence number of features as its life cycle and present a objective function weight setting method that adaptively considers the life cycle of extracted features. In pose estimation, the life cycle of feature can reflect the stability of features and the objective function weight based on it can decrease the accumulative error. We evaluate the proposed method on publicly available KITTI dataset. The experimental results demonstrate that the proposed method can achieve high-accuracy real-time visual odometry calculation.