GP-SLAM:Novel Laser-Based Simultaneous Localization And Mapping Method

One of the key issues for mobile robots to achieve truly autonomy is simultaneous local-ization and mapping(SLAM),which,for nearly thirty years,has been a research topic of great interest.The implementations of SLAM could rely on a variety of sensors.Among many types of them,the laser-based SLAM approach is widely-used owning to its high accuracy,even in poor lighting conditions.The laser range finder provide a large amount of unevenly distributed points described by noisy coordinates.In laser-based SLAM,the raw laser data are often transformed into a different map representation or model.Limited by the map representations,existing laser-based SLAM method exhibit limitations or drawbacks with regard to either efficiency or their map representations.Thus,this dissertation aim to create a new SLAM method.An ideal method should be accurate and fast,while provide a type of dense mapping results.It also should be uni-versal to two and three dimensions.The main contributions of the dissertation are summarized as follows:By reviewing the existing laser-based SLAM approaches,this dissertation believe that it is infeasible to make a breakthrough based on the existing map representations.Therefore,a new type of map representation is firstly proposed based on the regionalized Gaussian process(GP)map reconstruction algorithm,which is named as the GP map.The new type of map representation is compact and dense,its structure is between the raw point set and the traditional grid map which fully reflects the structure of the environment,and its memory consumption is low.Through the discussion and analysis of the definition of SLAM problem,this dissertation further set the research goal on redesigning the two core elements common to all SLAM systems,namely state estimation and map construction,based on the new type of map representation.By analyzing the structure of GP map,the iterative GP point set registration algorithm is proposed to solve the state estimation problem? by assuming each predictive point in the GP map to be independent,a map update method based on the recursive least-squares estimation is proposed.With this two algorithms,both the state estimation problem and the map update problem can be solved with simple and analytical mathematics.Meanwhile,the iterative GP point set registration algorithm can also be used as an independent scan-to-scan point set registration method.Finally,this dissertation combines the proposed state estimation method and map update method,and propose a new SLAM method with scan-to-map(St M)registration as its core the-oretical framework,which is named as GP-SLAM.Meanwhile,in order to verify the theory of St M extension,and to make GP-SLAM to be able to cope with larger scale environments,the dissertation extend our approach to a graph-based version.The proposed method is implemented and tested with a large number of experiments based on dataset measured from real-world envi-ronments.The experiment results indicate the GP-SLAM method proposed in this dissertation ex-hibits better performance with regard to efficiency,accuracy and memory consumption compared to the state-of-art approaches,and the method also has the universality between two dimensions and three dimensions.