动态环境下融合边缘信息的稠密视觉里程计算法

Dense visual odometry based on edge information fusion in dynamic environment

摘要:

针对传统的视觉里程计算法在动态环境下存在位姿估计精度不高且鲁棒性较差的问题,提出一种融合边缘信息的稠密视觉里程计算法.首先,使用深度信息计算像素点的空间坐标,并采用K-means算法进行场景聚类.分别基于光度信息与边缘信息的聚类构建出光度及几何一致性误差与边缘对齐误差,两者结合并进行正则化后得到数据融合的残差模型.将平均背景深度引入到残差模型中,用以扩大动、静部分残差差距而有利于正确的运动分割.然后,根据聚类残差分布的普遍特征,构建运动似然的非参数统计模型,通过动态阈值进行运动分割,剔除动态物体并得到聚类权重.最后,将加权聚类残差加入到位姿估计的非线性优化函数中,以降低动态物体的影响,提高位姿估计的精度.在TUM数据集上进行实验,结果表明本文算法在静态环境下以及富有挑战性的高动态环境下都能取得较好的结果,在动态环境下比现有算法具有更高的精度与鲁棒性.

Abstract:

In view of the low robustness and poor accuracy of traditional visual odometry in dynamic environment, a dense visual odometry based on edge information fusion was proposed. First, the spatial coordinates of pixels based on depth information were calculated, and the K-means algorithm was adopted for scene clustering. Based on the clustering of photometric information and edge information, the photometric consistency error and edge alignment error were constructed respectively, and the residual model was obtained after fusion and regularization of the two errors. Next, the average background depth was introduced to the residual model so as to expand the residual difference between the dynamic and static parts, ensuring correct motion segmentation. Then, a non-parametric statistical model was constructed based on the general characteristics of the cluster residual distribution, and motion segmentation was performed through dynamic thresholds to eliminate dynamic objects and obtain clustering weights. Finally, the weighted cluster residuals were added to the nonlinear optimization function of pose estimation to reduce the effect of dynamic objects and improve the accuracy of pose estimation. Experiments on TUM dataset show that the proposed algorithm could achieve better results in both static and high dynamic environments, and it had higher accuracy and robustness than the existing algorithm in dynamic environment.