Research On Motion Field Estimation And Applications

The extraction of motion information from images is one of the key technologys in the field of road scene perception of UGV.Motion field estimation based on images is widely researched in the applications ranging from dynamic target detection to highprecision map building to ego-motion estimation.Thus,in this article with the background of environment perception of UGV,different problems were solved by the same modelbased motion field estimation approach framework.Motion field estimation approach based on sequential images and multi-modality images was primarily researched.The main contributions and innovation are as follows:1)General framework of model-based motion estimation algorithm was sumed up.All steps including image data analysis,the selection of motion field model,similarity measure method,image warping and interpolation,objective function constructing and optimization,and motion field computetion were researched,and possible technical solutions of every step were given.The parameters of motion model were computed using Gaussian Newton's Method with inverse search strategy which is more efficient than the original version.2)A motion field estimaton approach based on multi-modality images was proposed.Firstly the concept of the similarity space was introduced in this approach,and patchbased stragety was used to complete the computation of the similarity spaces.Then a UGM-based rough motion field estimation method based on the local similarity spaces was proposed.Taking advantage of the rough motion field estimation result,a set of model parameters was fitted,which can be used as an initial value.Finally,global optimal model parameters,according to which the high-precision motion field could be computed,were optimized by minimizing the matching error.Experiment results show that proposed approach ensure the accuracy of each patch's motion field and the smooth transition between neighboring patches.3)A plannar flow model based motion field estimation approach from sequential images was represented.Aimed at camera egomotion estimation problem,an egomotion estimation algorithm based on monocular sequential images was achieved.Its effectiveness was verified by experiments.And an improved stereo sequential images based egomotion estimation approach was proposed.In this approach,Gaussian Process regression is used to detect the road area.The weight of pixels belong to road area was larger while matching the contiguous frames base on planar surface flow motion model.Experiment results show an improvement on the monocular version.