关键词: 光流场, 目标检测, 无人机, 计算机视觉, 非刚性块匹配, 大位移运动, 最近邻域场, 深度学习

Abstract: Optical flow field is an important basis for many computer vision tasks such as target detection and unmanned aerial vehicle positioning.In order to develop the accuracy and robustness of optical flow estimation suffered from the difficult motion such as non-rigid movement and large displacement motion,this paper proposes a large displacement optical flow estimation approach based on non-rigid dense patch matching.Firstly,we utilize the non-rigid dense patch matching to compute the initial nearest neighbor field between the consecutive frames,and eliminate the inconsistent regions of the computed nearest neighbor field according to the consistency of the neighboring patches in the image to obtain an accurate image nearest neighbor field.Secondly,we merge the nearest neighbor field into the TV-L1(Total Variational with L1 norm,TV-L1) optical flow model,and employ the nearest neighbor field to compensate the large displacement optical flow of TV-L1 model by using the quadratic pseudo-Boolean optimization (QPBO) fusion algorithm during the coarse-to-fine computation scheme.Finally,we employ the standard test image sequences to evaluate the performance of our approach and some state-of-the-art methods including LDOF(Large Displacement Optical Flow,LDOF),Classic+NL,NNF(Nearest Neighbor Fields,NNF) and FlowNet2.0.The experimental results demonstrate that the proposed method has the higher accuracy and better robustness of optical flow estimation for difficult motion such as non-rigid movement,large displacement motion and motion occlusion.

Key words: optical flow field, target detection, unmanned aerial vehicle, computer vision, non-rigid patch matching, large displacement motion, nearest-neighbor filed, deep learning