结合光流跟踪和三焦点张量约束的双目视觉里程计  

Stereo visual odometry based on Kalman fusion of optical flow tracking and trifocal tensor constraint

作　　者：程传奇[1,2] 郝向阳 张振杰[1,2] 赵漫丹 CHENG Chuan-qi HAO Xiang-yang ZHANG Zhen-jie ZHAO Man-dan(School of Navigation and Aerospace Engineering, Information Engineering University, Zhengzhou 450001, China Beidou Navigation Technology Collaborative Innovation Center of Henan, Zhengzhou 450001, China)

机构地区：[1]信息工程大学导航与空天目标工程学院,郑州450001 [2]北斗导航应用技术河南省协同创新中心,郑州450001

出　　处：《中国惯性技术学报》2016年第4期473-479,共7页Journal of Chinese Inertial Technology

基　　金：国家863计划资助项目(2015AA7034057A);国家自然科学基金项目(61173077)

摘　　要：针对城市街道准确实时定位的问题,提出将光流跟踪与三焦点张量约束结合的双目视觉里程计方法。为提高运算效率,将图像序列分为关键帧与非关键帧,对关键帧进行常规的特征点检测与匹配,对非关键帧用Lucas-Kanade光流跟踪特征点对。推导了基于前后帧、左右视图三焦点张量约束的观测方程,顾及动力学方程,组成卡尔曼滤波模型。考虑到观测方程的非线性,采用迭代Sigma点卡尔曼滤波进行解算,解算过程中用RANSAC稳健估计策略提纯匹配,以增强系统整体稳健性。实验结果表明:提出的算法在基本不损失精度(X方向优于5 m,Y方向优于4 m)的情况下,计算速度提升6.2倍,单帧图像平均处理时间由0.3115 s下降为0.0503 s,能够满足城市定位实时准确的需求。For the problem of real-time precise localization in the urban street, a stereo visual odometry based on Kalman fusion of optical-flow tracking and trifocal tensor constraint is proposed. To speed up the algorithm, the image sequences are classified into key frames and non-key frames. A conventional featurepoint detection and matching method is utilized to process the key frames, and the Lucas Kanade optical flow method is utilized to track the correspondences in non-key frames. The mathematical model of observation based on trifocal tensor constraint among image triples is derived, which is combined with dynamic equation to form the Kalman filter model. An iterated sigma-point Kalman filter is employed to cope with the nonlinear system. To improve the robustness of the visual odometry, a RANSAC-based method is applied into the pure matching points in motion estimation. Experimental results demonstrate that, compared with Kitt＇s method, the proposed method can acquire the same accuracy with the speed increasing 6.2 times, i.e., the error in the X direction and Y direction is 5 m and 4 m, respectively, while the average runtime of single frame decreases from 0.3115 s to 0.0503 s, which meets the needs of real-time accurate localization in cities.

关 键 词：双目视觉里程计 光流 三焦点张量 迭代Sigma点卡尔曼滤波 

分 类 号：V448.2[航空宇航科学与技术—飞行器设计]