Research On The Visual Mileage Calculation Method Of Unmanned Vehicles

With the computer technology and applications growing rapidly,tremendous changes have taken place to people's lifestyle,and smart devices are constantly replacing tradition labor and even traditional mechanical equipment.As one of the trends in the future,driverless vehicle has been a hotspot in a long time.Although many tech corporation,such as Google and Baidu,and some of automobile institute that established by traditional vehicle corporation,such as Mercedes-Benz and Volkswagen,have started test on the road,the existing automobile technology is still defective.Locating,the most basic technology,does have many limitation,thus the visual positioning technology,which as known as visual odometer technology,is a nice supplement.This thesis studies the shortage of the algorithm of visual positioning technology,proposes several strategies to prove the algorithm and proves it through the simulations.This algorithm relies on the feature points extracted from the image,and the quantity and discreteness of the feature points affect accuracy of the algorithm greatly.The thesis proposes a feature extracted algorithm,which based on splitting image.In this algorithm,firstly,split the image,and then extract each image fragment through the strategy by using adaptive weights.It is proved that the algorithm makes the feature discretely distribute over the whole image,and the quantity of feature could be constant while the features are extracted by different frames in the meantime.The algorithm uses the relationship among the different visual info by observing a scene at the different position,to solve the gesture changes from one position to another one.In case of the motive object affect the location accuracy,this thesis proposes that a motion estimation algorithm based weighted feature point that predicts position of current 3D feature points by testing the gesture changes of frame sequence,then contrast the estimate position and predict position,estimate the 3D feature point quality and contrast a motion estimate algorithm.It could test the motive objects in the scene and exclude most of violent object,finally makes the algorithm have high locate accuracy and robustness.In order to test the motive object in the scene,this thesis researches the traditional algorithm based on mixed-Gaussian motive object and proposes an improved strategy.The traditional algorithm is subject to be interfered by noise in practical use,and it costs much.The improved strategy is a mixed-Gaussian model algorithm based on splitting image.Using different mixed-Gaussian model due to different ratio of motive point,the algorithm pretreat those image fragment,which is at the edge of motive object.This measure not only decreases the influence of noise in background and motion area,but makes testing motive object more efficient.The result proves that the algorithm is feasible and robust.