Research On Binocular Camera Based Localization Algorithm Using Filtering And Machine Learning Technique

With the popularity of image sensors and the development of image recognition technology,image based localization technology has become a research hotspot.In order to meet high localization performance requirement in complex environment,In this thesis,binocular camera based localization algorithm using machine learning and filtering technology is studied.It can make full use of the provided grayscale and depth images with low hardware and time cost.Moreover,using the correlation between positions in continuous time,the target motion model is constructed to correct the positioning results and improve the localization performance.The main contributions are described as follows:1.The theoretical principle of image localization is studied and the hardware experiment platform is built.After studying the image feature extraction preprocessing and machine learning algorithms,the machine learning based image localization system model is described.Then,a hardware experiment platform based on Xiaomi binocular camera is built to collect grayscale and depth images which can provide a hardware basis for performance analysis of proposed algorithm2.A binocular camera based localization algorithm using support vector machine(SVM)and Singer motion model is proposed.The problem of localization is formulated as a theoretical image regression learning based machine learning model at first.In the offline phase,the local binary mode(LBP)features and depth information are extracted from the grayscale image and depth image respectively,and then the above feature information are used for regression learned by SVM respectively.The regression functions based on X axis and Y axis are obtained at last.In the online phase,the extracted image feature information and regression function are used for coarse localization.Then the Singer motion model is used to correct the positioning results and can achieve precise localization.Experimental results show that the proposed algorithm has better localization performance than the existing localization algorithms.3.A binocular camera based localization algorithm using SVM and interactive multi-model(IMM)is proposed.The target motion model selection in IMM is studied at first.And then the filtering process of IMM is introduced.Next,the steps of the proposed algorithm are described in detail.Compared with a single motion model,the proposed algorithm has a better filtering effect because the precision localization result is obtained with the weighted outputs of each model.Experimental results show the effectiveness of the proposed algorithm.