URBAN-NAV: Urban-Navigation of Unmanned Platform under the GPS Challenged Environments
Tuesday, 28 February 2017 10:10



urban-nav1 urban-nav2

This project contains three main topics: “Location Estimation Using Panoramic View Vision”, “Pose Estimation and Localisation Method Using Visual Odometry”, and “Non-GPS Localization Using Local Geometrical Constraints”. Among them we mainly focus on the third approach. The constrained solution is proposed by approximately modeling the path of the vehicle in the urban canyon environment as pieces of curve. By applying these constraints and implementing multiple sensor data such as visual odometry and inertial measurement unit, the necessary number of GPS satellite can be waived. The challenges, risks and methodologies are described below.

Road Modeling. Fortunately, detailed maps which can be modeled as junctions connected by piecewise continuous curves are usually available for most cities. The model of the road on which the vehicle is traveling can be extracted and regarded as constraints to facilitate the positioning with no GPS signals in view. Visual odometry and local simultaneous localization and mapping (“SLAM”) can be used to assist positioning the vehicle.

Monte-Carlo Localization. Based on the proposed measurement and motion models, a set of particles can be generated from the error model. The particle set covers all probable trajectories of the vehicle, and each particle in the set is assigned an initial weight according to the raw measurement. For real implementation, the place recognition results can be used in initialization process and in environments without drivable roads. This topological localization method, together with odometry measurement is integrated to the Monte-Carlo localization frame for better solution. For better real-time performance, the parameter estimation step is running in parallel within positioning to limit parameter range.

Shape Matching. In order to measure the similarity between road model and probable vehicle trajectories, shape matching techniques in computer vision can be applied. The trajectories with best matching similarities will be used to estimate the true position of the vehicle. At the same time, measurement error of other sensors can be decreased after shape matching process.