结合路网地图的视觉定位优化方法研究

发布时间：2018-05-14 07:52

本文选题：视觉定位 + 视觉里程计　；参考：《浙江大学》2017年硕士论文

【摘要】：移动机器人的自定位问题是机器人领域的一个关键问题。摄像头作为移动机器人的“眼睛”,由于其体积小、成本低、应用场景广的特点而得到了广泛的应用。由于传统的定位手段比如GPS、惯导在城市、室内环境下定位结果不稳定,近些年来,视觉定位越来越受到了广泛的关注。视觉里程计是视觉定位中一种经典的方法,但是,视觉里程计存在累积误差的问题,无法在实际的长距离中进行运用。在本文中,我们提出了一种新颖的基于多位置联合滤波和路网地图的定位算法来解决视觉里程计的误差累积问题。与以往的基于地图的定位方法不同,我们认为基于“点-线”的地图表示方法并不能很精确地表示地图,我们的方法充分考虑到地图的这种不精确性,不会强制将滤波后的车辆轨迹纠正到地图上。将视觉里程计定位结果和路网地图信息作为初始输入,在多位置联合粒子滤波框架下设计了一种灵活的定位优化算法。该算法将视觉里程计的定位结果作为初步的轨迹与路网地图进行匹配滤波,使定位结果能在视觉里程计定位和路网地图匹配定位中进行合适的平衡。由于只在车辆轨迹的拐点处进行滤波,因此相对于视觉里程计,算法只增加了很少的计算量。在KITTI数据集和校园环境中采集的数据进行了多个实验,并与其他定位算法进行了定量的比较,实验结果都表明了所提算法的准确性和鲁棒性。
[Abstract]:The self-localization of mobile robot is a key problem in the field of robot. As the "eye" of mobile robot, camera is widely used because of its small size, low cost and wide application scene. Because of the instability of traditional positioning methods such as GPS, inertial navigation in cities and indoor environment, visual positioning has been paid more and more attention in recent years. Visual odometer is a classical method in visual positioning. However, there is a problem of accumulated error in visual odometer, which can not be used in actual long distance. In this paper, we propose a novel localization algorithm based on multi-position joint filtering and road map to solve the error accumulation problem of vision odometer. Different from the previous map based localization methods, we think that the map representation method based on "point-line" can not represent the map accurately, and our method fully takes into account the imprecision of the map. The filtered vehicle trajectory will not be forced to be corrected onto the map. The location results of vision odometer and road map information are taken as the initial input, and a flexible location optimization algorithm is designed under the framework of multi-position joint particle filter. The algorithm uses the location result of visual odometer as the initial path and the road map to match and filter, so that the location results can be properly balanced between vision odometer location and road map matching location. Due to filtering only at the inflection point of the vehicle trajectory, the algorithm increases only a small amount of computation compared to the visual mileometer. The data collected in KITTI data set and campus environment are tested and compared quantitatively with other localization algorithms. The experimental results show that the proposed algorithm is accurate and robust.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】