利用Kinect点云数据的室内未知环境自主定位

发布时间：2019-01-19 15:27

【摘要】：随着机器人定位研究的不断发展,同时定位测图(SLAM)已成为计算机视觉的一个研究热点。主要的研究内容是在未知的环境中,利用传感器获取的数据在实现场景地图生成的同时,完成定位。随着RGB-D相机的发布,为SLAM问题提供了一种新的传感器。Kinect作为RGB-D相机的一种,随着微软公司开发的Kinect不断的更新升级,利用Kinect设备实现未知环境定位也成为了一大研究热点。现阶段,利用Kinect相机进行定位的方法有以下几种:利用Kinect获取的数据通过迭代近邻点(ICP)实现定位。但是受到场景范围以及Kinect深度影像自身精度不高的影响,导致最后通过ICP算法进行配准的结果的误差累计比较大,定位精度比较低;利用SLAM的方法进行室内场景的定位及地图的生成,但是当场景不闭合或者闭合场景范围较大时,回环检测的处理后的结果也不是很好。针对以上情况,本文以ORBSLAM系统为基础,通过ORB特征提取的方式,对Kinect获取的序列影像中的彩色影像的特征点进行提取利用,实现影像的跟踪,获取相邻帧影像之间的相对转换关系,对先验信息向量机信息矩阵进行求解,完成多特征信息滤波中的信息向量更新及状态增广;针对深度影像,利用彩色影像和深度影像的标定结果,对序列影像的彩色点云数据进行生成,对生成的点云数据,完成场景内点特征及面特征的提取,依据提取的特征结果,完成对多特征观测模型的构建;依据观测模型以及状态增广的结果,对状态进行更新,求解后验信息矩阵及信息向量。削弱在定位过程中误差累计的影响,对更新后的状态向量中的平移参数进行求解,完成利用Kinect点云数据在未知环境中的自主定位。本文利用Kinect传感器,选择了2个实验场景进行了实验验证。实验内容包括:利用ORBSLAM系统模型对数据进行获取,实现影像跟踪;在此基础上,利用ICP获取的点特征以及场景中的面特征构建多特征扩展信息滤波模型,对实验数据进行处理。实验结果表明:利用本文提出的方法能够在很大程度上削弱误差累计对Kinect数据的影响,针对实验场景1,与融合数据进行比较,定位精度在0.2m范围内,针对实验场景2,用ORBSLAM的方法和本文方法分别实现定位,定位精度分别为0~0.8m和0~0.5m,定位中误差分别为0.48m和0.23m。
[Abstract]:With the development of robot localization research, simultaneous location mapping (SLAM) has become a hotspot in computer vision. The main research content is to complete the location of scene map while generating scene map by using the data obtained by sensor in unknown environment. With the release of the RGB-D camera, it provides a new sensor for the SLAM problem. As one of the RGB-D cameras, Kinect has been continuously updated and upgraded with the Kinect developed by Microsoft. Using Kinect equipment to realize unknown environment positioning has also become a research hotspot. At present, there are several methods for positioning by using Kinect camera: the data obtained by Kinect can be located by iterative nearest neighbor (ICP). However, due to the low precision of scene range and Kinect depth image, the result of registration through ICP algorithm has a large cumulative error and low positioning accuracy. The method of SLAM is used to locate the indoor scene and generate the map, but when the scene is not closed or the scope of the scene is large, the result of loop detection is not very good. In view of the above situation, based on the ORBSLAM system, this paper uses the ORB feature extraction method to extract and utilize the feature points of the color images obtained by Kinect, so as to realize the tracking of the images. The relative transformation relationship between adjacent frames is obtained and the information matrix of prior information vector machine is solved to update and expand the information vector in multi-feature information filtering. For the depth image, using the calibration results of color image and depth image, the color point cloud data of the sequence image is generated, and the point cloud data is extracted from the generated point cloud data. The construction of multi-feature observation model is completed. According to the observation model and the result of state expansion, the state is updated and the posterior information matrix and information vector are solved. By weakening the effect of error accumulation in the localization process, the translation parameters in the updated state vector are solved, and the autonomous localization in unknown environment using Kinect point cloud data is accomplished. In this paper, two experimental scenarios are selected and verified by using Kinect sensor. The experimental contents include: using the ORBSLAM system model to obtain the data and realize the image tracking; on this basis, using the point features acquired by ICP and the surface features in the scene to construct a multi-feature extended information filtering model to process the experimental data. The experimental results show that the method proposed in this paper can greatly weaken the effect of error accumulation on Kinect data. For experiment scenario 1, compared with fusion data, the positioning accuracy is within 0.2 m, and for experimental scenario 2. The method of ORBSLAM and the method of this paper are used to realize the positioning. The positioning accuracy is 0 0. 8 m and 0 0. 5 m, and the error in the positioning is 0. 48 m and 0. 23 m, respectively.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212;TP391.41

【相似文献】