Research On Intelligent Robot Local Visual Homing Algorithm Based On Natural Landmarks

Visual navigation becomes one of the important subjects that deserve intensive research in the field of intelligent robot,because it can attain resourceful and adequate environment information.At present,the vision-based navigation methods mainly fall into two categories: one is called SLAM(Simultaneous Localization and Mapping),and it is based on localization.Methods of this kind need constantly update the map of the environment and position the robot itself during the process of navigation,so a large amount of calculation has to be conducted and the problem of accumulative error has to be solved.The other one is referred to as visual homing;it is advantageous for its saving computing resources and non accumulative error when compared to SLAM.Visual homing is able to solve the problems of large-scale navigation solely by combing topological map of the environment.So this dissertation conducts an intensive research on local visual homing based on natural landmarks.First of all,the experiments are designed on the basis of image database and real scene to test the performance of different homing methods,which in turn lay a foundation for the subsequent comparative homing experiments.The discrete coordinate corresponded to image database is built by analyzing the characteristic of panoramic databases.In order to test the methods in real scene,a mobile robot platform based on catadioptric panoramic imaging system is built up and a method to remove the invalid field in panoramic images is proposed.On the basis of all the above,this dissertation designs a corresponding panoramic coordinate system and a world coordinate system,analyzes the four criteria of homing angular deviation,average homing angular deviation,average homeward component and return ratio,and discusses the parameter setting problem in return ratio experiment to evaluate the performance of each method quantitatively.Secondly,the research focuses on the extraction,selection and optimization of natural landmarks.Although favorable characteristics as easy identification and high matching precision are reserved in artificial landmarks,it is sometimes not convenient to set up landmarks beforehand in practical application.Under such circumstances,it is necessary for the robot to extract some local features like edge,corner,spot and special areas as natural landmarks to fulfill its navigation task.For the sake of its great stability,the local features of SIFT(Scale Invariant Feature Transform)is extensively employed as visual landmarks during this process,but the proof for its rationale still remains to be seen.For the above reasons,the focus of this research is on the extraction of SIFT features and the proof for the rationale of selecting SIFT features as natural landmarks according to the following three principles when selecting landmarks: uniqueness,reliability and correlation.The homing methods proposed in this dissertation all adopt SIFT features as landmarks,so the homing performance can be improved by eliminating mismatching landmark pairs.For this reason,two distribution constraints are given by studying the catadioptric panoramic imaging characteristics,then two mismatching elimination algorithms are proposed on the basis of landmarks' distribution and azimuth variation characteristics.Based on the above,this dissertation studies two representative homing methods namely ALV(average landmark vector)and warping.The two methods are separately based on parameter model and snapshot model.The advantage of ALV is its simpler model and less storage space.For warping,its homing accuracy is higher and its performance is more stable in spite of its large amount of calculation.In the end,the optimization of the above two homing methods is intensively studied.On the one hand,to the problem that ALV method extracting landmarks within the whole image leads to a great number of landmarks and thus makes it impossible to ensure their correspondence,a new concept,horizon ring-shaped region,is proposed based on the characteristics of catadioptric panoramic imaging.It is in this ring-shaped region that SIFT features as natural landmarks are extracted and furthermore an improved ALV algorithm is proposed with the combination of the extracted landmarks and ALV model.The results of homing experiments that are based on image databases suggest that the improved ALV algorithm effectively reduces the number of landmarks,ensures the correspondence of landmarks and further improves the homing accuracy.On the other hand,in order to solve that the homing performance of warping is greatly influenced by environment variation,an improved warping method based on the natural landmarks is proposed by adopting SIFT features as landmarks instead of pixels on horizon circle.The results of homing experiments based on databases and real scene indicate that the improved warping method is advantageous than original one from the point of overall performance according to the above four criteria and that it is more robust to the variation of objects and illumination of the surroundings.