外文翻译-一个有关移动机器人定位的视觉传感器模型.doc

1、毕业设计(论文)外文资料翻译院 系专 业学生姓名班级学号外文出处Machine Vision and Applications指导教师评语 指导老师签名： 日期A Visual-Sensor Model for Mobile Robot LocalisationMatthias Fichtner Axel Gro_mannArti_cial Intelligence InstituteDepartment of Computer ScienceTechnische Universitat DresdenTechnical Report WV-03-03/CL-2003-02AbstractWe

2、 present a probabilistic sensor model for camera-pose estimation in hallways and cluttered o_ce environments. The model is based on the comparison of features obtained from a given 3D geometrical model of the environment with features presentin the camera image. The techniques involved are simpler t

3、han state-of-the-art photogrammetric approaches. This allows the model to be used in probabilistic robot localisation methods. Moreover, it is very well suited for sensor fusion. The sensor model has been used with Monte-Carlo localisation to track the position of a mobile robot in a hallway navigat

4、ion task. Empirical results are presented for this application.1 IntroductionThe problem of accurate localisation is fundamental to mobile robotics. To solve complex tasks successfully, an autonomous mobile robot has to estimate its current pose correctly and reliably. The choice of the localization

5、 method generally depends on the kind and number of sensors, the prior knowledge about the operating environment, and the computing resources available. Recently, vision-based navigation techniques have become increasingly popular 3. Among the techniques for indoor robots, we can distinguish methods

6、 that were developed in the _eld of photogrammetry and computer vision, and methods that have their origin in AI robotics.An important technical contribution to the development of vision-based navigationtechniques was the work by 10 on the recognition of 3D-objects from unknown viewpoints in single

7、images using scale-invariant features. Later, this technique was extended to global localisation and simultaneous map building 11.The FINALE system 8 performed position tracking by using a geometrical model of the environment and a statistical model of uncertainty in the robots pose given the comman

8、ded motion. The robots position is represented by a Gaussian distribution and updated by Kalman _ltering. The search for corresponding features in camera image and world model is optimized by projecting the pose uncertainty into the camera image.Monte Carlo localisation (MCL) based on the condensati

9、on algorithm has been applied successfully to tour-guide robots 1. This vision-based Bayesian _ltering technique uses a sampling-based density representation. In contrast to FINALE, it can represent multi-modal probability distributions. Given a visual map of the ceiling, it localises the robot glob

10、ally using a scalar brightness measure. 4 presented a vision-based MCL approach that combines visual distance features and visual landmarks in a RoboCup application. As their approach depends on arti_cial landmarks, it is not applicable in o_ce environments.The aim of our work is to develop a probab

11、ilistic sensor model for camerapose estimation. Given a 3D geometrical map of the environment, we want to find an approximate measure of the probability that the current camera image has been obtained at a certain place in the robots operating environment. We use this sensor model with MCL to track

12、the position of a mobile robot navigating in a hallway. Possibly, it can be used also for localization in cluttered o_ce environments and for shape-based object detection.On the one hand, we combine photogrammetric techniques for map-based feature projection with the exibility and robustness of MCL,

13、 such as the capability to deal with localisation ambiguities. On the other hand, the feature matching operation should be su_ciently fast to allow sensor fusion. In addition to the visual input, we want to use the distance readings obtained from sonars and laser to improve localisation accuracy.The

14、 paper is organised as follows. In Section 2, we discuss previous work. In Section 3, we describe the components of the visual sensor model. In Section 4, we present experimental results for position tracking using MCL. We conclude in Section 5.2 Related WorkIn classical approaches to model-based po

15、se determination, we can distinguish two interrelated problems. The correspondence problem is concerned with _nding pairs of corresponding model and image features. Before this mapping takes place, the model features are generated from the world model using a given camera pose. Features are said to match if they are located close to each other. Whereas the pose problem consists of _nding the 3D camera coordinates with respect to the origin of the world model given the pairs of corresponding features 2. Apparently, the one problem requires the other to be solved befo