Applying geometric sensor and scene models for range image understanding

摘要

Abstract: For a robot to operate intelligently under sensorcontrol in the real world, it must interpret itssensory inputs to create a model of its environment.Because the data is generally incomplete and errorful,additional knowledge must be applied to derive usefulinterpretations. One source of such data is knowledgeabout the sensors themselves; the ideal data returnedfrom a sensor is a function of the environment and thepose and other parameters of the sensor (such as thefocal length of a camera). The actual result is somecorruption of the ideal result by noise or other errorprocesses. In this work, the functions 'computed' by arange sensor are explicitly represented as geometricobjects and relationships in the 3D FORM geometricreasoning system. Models for the output of independentedge segmentation and surface segmentation aredescribed, and unified into an overall range-objectmodel. In addition, visibility and projectionrelationships are used to predict the visibility ofhypothesized object parts, and to refine the estimatesof the cameras' positions. The resulting objectdescription can be specialized, completed, and matchedwith other objects using existing 3D FORMcapabilities.!20