基于动态物体特征点去除的视觉里程计算法

Visual Odometry Algorithm Based on Dynamic Object Feature Point Removal

摘要:

存在运动物体的场景中,传统立体视觉里程计定位误差大,系统鲁棒性低。为了有效解决此问题,提出了一种基于动态物体检测与特征点去除的立体视觉里程计算法。利用二维稀疏光流法计算由相机运动带来的运动变换矩阵,然后利用变换矩阵对相机运动进行补偿。进一步地,采用了向前向后帧间差分方法进行运动物体检测,将场景中的动态物体和静态物体初步区分开来。将检测到的运动物体区域作为mask,去除动态物体上的部分特征点,然后利用静态点集继续优化变换矩阵,将动态物体影响不断减小,并将其应用于立体视觉里程计系统中。经TUM的RGBD数据集测试评估,提出的算法有效提高了动态场景下的视觉里程计定位精度。

Abstract:

In the scene with moving objects, the traditional stereo visual odometry has large positioning error and low system robustness. In order to solve this problem effectively, a stereo visual odometry algorithm based on moving object detection and feature point removal is proposed. The motion transformation matrix brought by the camera motion is calculated by the two-dimensional sparse optical flow method, and then the motion of the camera is compensated by the transformation matrix. Further, the forward-backward inter-frame difference algorithm is used for moving object detection, and the dynamic object and the static object in the scene are initially distinguished. Using the detected moving object’s area in the image as a mask, some feature points on the dynamic object are removed, and then the static point set is used to continue to optimize the transformation matrix, and the dynamic object influence is continuously reduced, and is applied to the stereo vision odometry system. After TUM"s RGBD dataset test evaluation, the proposed algorithm effectively improves the visual odometry positioning accuracy under dynamic scene.