Abstract: Tracking features on image time series have huge potential for many applications, but the popular optical flow tracking algorithm is less practical because of its less accuracy. To tackle this problem, this paper proposes a method of tracking endpoints of indoor structure lines based on backprobing of optical flow. The endpoints of the building structure lines indoor are used as target features to be detected on the reference frame of image series, then the features are tracked using the optical flow algorithm to obtain the corresponding positions on the next frame of images; The idea is backprobing which means tracking the corresponding positions back to the reference frame to get the backprobing positions, and comparison is made between the original feature positions and the backprobing positions which should be no difference ideally on the reference frame; The effective features are finally selected by eliminating the ones with large differences. This improved optical flow tracking algorithm is applied to the visual odometry, and the experiment results show that the image features of structure-line-endpoints are stable, and the backprobing optical flow tracking method has a high accuracy, and the final trajectory of the visual odometry is significantly improved compared with the traditional optical flow algorithm.

Key words: Optical flow, feature tracking, backprobing, visual odometry, image feature detection

摘要： 时间序列图像的特征跟踪技术有着巨大的潜在应用，然而传统的光流法跟踪算法由于其精度不高而影响了其实用性。针对这一问题提出一种基于回测光流法的结构线端点跟踪方法。将图像上的建筑结构线的端点作为目标特征，在序列图像的当前参考帧进行特征检测，利用光流跟踪法得到特征在下一帧图像上的对应位置；进一步将跟踪到的特征位置点用于回测跟踪，将回测跟踪的结果与参考帧原始检测到的特征位置进行比较，排除掉误差大的特征点,最终获得有效特征检测结果。将改进后的光流法应用于视觉里程计，实验结果表明，结构线端点特征稳定，回测光流跟踪精度较高，视觉里程计的计算轨迹精度比应用传统光流法有明显提高。

关键词: 流法, 特征跟踪, 回测, 视觉里程计, 图像特征检测