Hybrid Map Based Localization Method For Unmanned Ground Vehicle In Urban Scenario

This thesis is focus on the map matching based localization method for unmannedintelligent vehicle in urban scenarios, to solve the positioning accuracy problems caused bycumulative error and obscured satellite signals of traditional localization methods. Toimprove the positioning precision of map matching positioning affected by the accuracy ofenvironment map and the realtime performance affected by the scale of the map, a hybridmap based localization method for intelligent vehicle was proposed in this thesis. Thecontent can be divided into three parts: the research of creating hybrid map, the hybrid mapbased localization method for intelligent vehicle and experiments based on simulationsystem and vehicle platform.Firstly, the structure of hybrid map constructed by topological and metric map wasanalyzed. By dividing the urban scenario into Strong and Loose Constraint Areas, thetopological map was built by defining Loose Constraint Areas as nodes of the topologicalmap, which were connected by Strong Constraint Areas as edges. Occupancy grid mapscreated by Simultaneous Localization and Mapping method were used to provide additionalrestrictions to Loose Constraint Area. The topological map was constructed using satellitepositioning data and its nodes were associated with landmarks built in the grid map toextract local metric map around corresponding node. Experimental results showed that itgreatly reduced the data size of map used during autonomous driving. In the process of gridmap construction, a sub map based Simultaneous Localization and Mapping method wasproposed. Experimental results showed that this method could build the map incrementallyand guarantee the switching process smoothly and unidirectionally between sub maps.Secondly, based on metric-topological hybrid map, the relative positioning strategyusing environment elements to constrain vehicle movement was analyzed. In the StrongConstraint Areas, a multi-feature extraction method for drivable road region detection usingLaser Range Finder was proposed. Curbs were extracted from the range data using thewavelet transform method and divide the data into pieces. Considering characteristics suchas length, flatness and consistency of echo data, drivable road regions were detected.Experimental results verified the feasibility and stability of the method. In the LooseConstraint Areas, using local occupancy grid map registered to the node, a windowconstrained Markov localization method was proposed. The scope of state space was constrained by doing Markov localization within the window. This method reduced theamount of calculation and did not rely on satellite positioning system besides the initialtime.In the end, the2D simulation system and vehicle platforms for urban autonomousdriving were developed. For the simulation system, a multi-layer map was used as theenvironment model of the simulation platform to express the static obstacles, movingobjects and traffic lanes. In addition, the localization module, distance measurementmodule and lane detection module were designed. The results of experiment showed thatthis system could describe the scene of structured environment well. The design ofintelligent vehicle for urban autonomous driving was also proposed in the end. The windowconstraint Markov Localization method was tested based on both simulation system and thevehicle platform to prove its positioning accuracy and realtime performance.