基于三维激光雷达的动态车辆检测与跟踪
发布时间：2019-06-13 14:28
【摘要】：物体检测与跟踪是移动机器人领域中的核心问题之一。作为物体检测与跟踪的重要组成部分,动态车辆检测与跟踪对自主车环境感知能力的提高有重要作用。本文以自主车在城市环境中自主导航为背景,重点研究了基于三维激光雷达的动态车辆检测与跟踪问题。论文的主要成果和创新点如下:(1)提出了两种新的地面分割算法对各种地形的三维激光雷达数据进行分割,分别是:基于区域高斯过程回归的地面分割算法和基于分块递归高斯过程回归的实时地面分割算法。基于区域高斯过程回归的地面分割算法主要应用于三维笛卡尔坐标栅格地图,该算法使用带有稀疏协方差函数的二维高斯过程回归直接对所有栅格单元中高度最低的三维点进行地面建模。在公开的波士顿数据库上,该算法地面分割的准确率达到97.90%。基于分块递归高斯过程回归的实时地面分割算法将在三维笛卡尔坐标栅格地图中复杂的、大尺度的二维地面分割问题分解为极坐标栅格地图中的多个低复杂度的一维回归问题,对于极坐标栅格地图的每个扇形块,分别采用带有非静态协方差函数的一维递归高斯过程回归算法对相应区域内的局部地面进行建模。在同样的数据库上该算法的准确率为97.67%,同时还能够满足自主车所必须的实时性要求。(2)提出了一种基于迭代高斯过程回归的道路边界检测算法来获取自主车的感兴趣区域。该算法以三维激光雷达的每一条扫描线为处理单元提取其中的特征点,然后采用迭代高斯过程回归算法根据提取的特征点自动对直线或曲线道路边界进行建模。本文提出的道路边界检测算法在满足检测精度的情况下能够检测到离自主车50米远的道路边界。为了定量地验证该道路边界检测算法的性能,本文手工标记了一个基于三维激光雷达的道路边界数据库。在该数据库上使用本文提出的基于迭代高斯过程回归的道路边界检测算法,左右道路边界的正检率分别达到 78.74%和 81.96%。(3)提出了一种新的全局柱坐标直方图特征用于在城市环境进行车辆识别。该特征以感兴趣区域内每个物体的中心为原点,通过引入全局坐标系来克服物体绕z轴旋转的不变性,并将柱形支持域内的所有三维点按其柱坐标进行划分,构建三维直方图。在悉尼城市物体数据库和我们整理、标记的数据集上进行车辆识别的ROC曲线验证了新的全局柱坐标直方图特征在车辆识别上的优异性能。(4)提出了一种新的基于似然场模型的动态车辆检测与跟踪算法。该算法首先使用我们提出的基于似然场的车辆观测模型结合改进的Scaling Series算法来估计感兴趣区域内各个车辆的初始姿态。在动态车辆检测阶段,本文改进了一种基于二维虚拟帧的三维激光雷达数据表示方式,采用该表示方式的动态车辆检测算法能够检测到感兴趣区域内在xy平面完全被其它物体遮挡,但仍能够被三维激光雷达感知到的动态车辆;在跟踪阶段,本文提出了一种新的基于贝叶斯滤波器的变尺寸车辆跟踪算法,由于引入了不动点,该跟踪算法不仅能在动态背景场景中更新目标车辆的姿态和速度,而且能够在跟踪过程中根据所关联的观测数据自动更新目标车辆的尺寸。在公开的KITTI数据库和在城市、高速公路环境采集的激光雷达数据上的定量和定性实验都验证了本文提出的动态车辆检测与跟踪算法的性能。以上研究成果已经成功应用于本实验室的自主车"开路雄师",该自主车在2015年第七届中国智能车"未来挑战赛"中获得了第三名。
[Abstract]:Object detection and tracking is one of the core problems in the field of mobile robots. As an important part of object detection and tracking, dynamic vehicle detection and tracking plays an important role in improving the environment-sensing ability of autonomous vehicle. This paper focuses on the dynamic vehicle detection and tracking problem based on three-dimensional lidar in the background of autonomous navigation in the urban environment. The main achievements and innovation points of the paper are as follows: (1) Two new ground segmentation algorithms are proposed to divide the three-dimensional lidar data of various terrain, respectively: A ground segmentation algorithm based on regional Gaussian process regression and a real-time ground segmentation algorithm based on block recursive Gaussian process regression. The terrain segmentation algorithm based on the regional Gaussian process regression is mainly applied to a three-dimensional Cartesian coordinate grid map, which uses a two-dimensional Gaussian process regression with a sparse covariance function to directly model the three-dimensional points with the lowest height in all the grid cells. The accuracy of the algorithm is 97.90% on the open Boston database. The real-time ground segmentation algorithm based on the block recursive Gaussian process regression is used for decomposing a complex and large-scale two-dimensional ground segmentation problem in a three-dimensional Cartesian coordinate grid map into a plurality of low-complexity one-dimensional regression problems in a polar coordinate grid map, For each sector of the polar coordinate grid map, a one-dimensional recursive Gaussian process regression algorithm with a non-static covariance function is used to model the local ground in the corresponding region. The accuracy of the algorithm is 97.67% on the same database, and the real-time requirement of the autonomous vehicle can be met. (2) A road boundary detection algorithm based on the iterative Gaussian process regression is proposed to obtain the region of interest of the autonomous vehicle. The method uses each scanning line of the three-dimensional laser radar as the processing unit to extract the characteristic points, and then uses the iterative Gaussian process regression algorithm to automatically model the straight line or the curve road boundary according to the extracted feature points. The road boundary detection algorithm proposed in this paper can detect the road boundary which is 50 meters away from the autonomous vehicle when the detection accuracy is satisfied. In order to quantitatively verify the performance of the road boundary detection algorithm, a road boundary database based on a three-dimensional laser radar is manually marked. The road boundary detection algorithm based on the iterative Gaussian process regression is used in the database, and the positive rate of the left and right road boundary is 78.74% and 81.96%, respectively. (3) A new global column coordinate histogram is proposed for vehicle identification in urban environment. The characteristic takes the center of each object in the region of interest as the origin, overcomes the invariance of the rotation of the object around the z-axis by introducing a global coordinate system, and divides all three-dimensional points in the cylindrical support domain according to the column coordinates to construct a three-dimensional histogram. The ROC curve of vehicle identification on the database of the city object in Sydney and the data set that we have sorted and marked verifies the excellent performance of the new global column coordinate histogram feature on vehicle identification. (4) A new dynamic vehicle detection and tracking algorithm based on the likelihood field model is proposed. The algorithm first uses the likelihood-based vehicle observation model proposed by us to combine the improved scaling series algorithm to estimate the initial attitude of each vehicle in the region of interest. in that dynamic vehicle detection stage, the invention improve a three-dimensional laser radar data representation mode based on a two-dimensional virtual frame, but still can be perceived by the three-dimensional laser radar; in the tracking phase, a novel variable-size vehicle tracking algorithm based on a Bayesian filter is proposed, The tracking algorithm can not only update the attitude and speed of the target vehicle in the dynamic background scene, but also can automatically update the size of the target vehicle according to the associated observation data during the tracking process. The performance of the dynamic vehicle detection and tracking algorithm presented in this paper is verified by the quantitative and qualitative experiments on the open KITTI database and the lidar data collected in the city and the highway environment. The above research results have been successfully applied to the autonomous vehicle "open-circuit male" of the laboratory, and the independent vehicle has obtained the third place in the first "future competition" of the 7th China Intelligent Vehicle in 2015.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：U495;TN958.98
，
本文编号：2498583
[Abstract]:Object detection and tracking is one of the core problems in the field of mobile robots. As an important part of object detection and tracking, dynamic vehicle detection and tracking plays an important role in improving the environment-sensing ability of autonomous vehicle. This paper focuses on the dynamic vehicle detection and tracking problem based on three-dimensional lidar in the background of autonomous navigation in the urban environment. The main achievements and innovation points of the paper are as follows: (1) Two new ground segmentation algorithms are proposed to divide the three-dimensional lidar data of various terrain, respectively: A ground segmentation algorithm based on regional Gaussian process regression and a real-time ground segmentation algorithm based on block recursive Gaussian process regression. The terrain segmentation algorithm based on the regional Gaussian process regression is mainly applied to a three-dimensional Cartesian coordinate grid map, which uses a two-dimensional Gaussian process regression with a sparse covariance function to directly model the three-dimensional points with the lowest height in all the grid cells. The accuracy of the algorithm is 97.90% on the open Boston database. The real-time ground segmentation algorithm based on the block recursive Gaussian process regression is used for decomposing a complex and large-scale two-dimensional ground segmentation problem in a three-dimensional Cartesian coordinate grid map into a plurality of low-complexity one-dimensional regression problems in a polar coordinate grid map, For each sector of the polar coordinate grid map, a one-dimensional recursive Gaussian process regression algorithm with a non-static covariance function is used to model the local ground in the corresponding region. The accuracy of the algorithm is 97.67% on the same database, and the real-time requirement of the autonomous vehicle can be met. (2) A road boundary detection algorithm based on the iterative Gaussian process regression is proposed to obtain the region of interest of the autonomous vehicle. The method uses each scanning line of the three-dimensional laser radar as the processing unit to extract the characteristic points, and then uses the iterative Gaussian process regression algorithm to automatically model the straight line or the curve road boundary according to the extracted feature points. The road boundary detection algorithm proposed in this paper can detect the road boundary which is 50 meters away from the autonomous vehicle when the detection accuracy is satisfied. In order to quantitatively verify the performance of the road boundary detection algorithm, a road boundary database based on a three-dimensional laser radar is manually marked. The road boundary detection algorithm based on the iterative Gaussian process regression is used in the database, and the positive rate of the left and right road boundary is 78.74% and 81.96%, respectively. (3) A new global column coordinate histogram is proposed for vehicle identification in urban environment. The characteristic takes the center of each object in the region of interest as the origin, overcomes the invariance of the rotation of the object around the z-axis by introducing a global coordinate system, and divides all three-dimensional points in the cylindrical support domain according to the column coordinates to construct a three-dimensional histogram. The ROC curve of vehicle identification on the database of the city object in Sydney and the data set that we have sorted and marked verifies the excellent performance of the new global column coordinate histogram feature on vehicle identification. (4) A new dynamic vehicle detection and tracking algorithm based on the likelihood field model is proposed. The algorithm first uses the likelihood-based vehicle observation model proposed by us to combine the improved scaling series algorithm to estimate the initial attitude of each vehicle in the region of interest. in that dynamic vehicle detection stage, the invention improve a three-dimensional laser radar data representation mode based on a two-dimensional virtual frame, but still can be perceived by the three-dimensional laser radar; in the tracking phase, a novel variable-size vehicle tracking algorithm based on a Bayesian filter is proposed, The tracking algorithm can not only update the attitude and speed of the target vehicle in the dynamic background scene, but also can automatically update the size of the target vehicle according to the associated observation data during the tracking process. The performance of the dynamic vehicle detection and tracking algorithm presented in this paper is verified by the quantitative and qualitative experiments on the open KITTI database and the lidar data collected in the city and the highway environment. The above research results have been successfully applied to the autonomous vehicle "open-circuit male" of the laboratory, and the independent vehicle has obtained the third place in the first "future competition" of the 7th China Intelligent Vehicle in 2015.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：U495;TN958.98
，
本文编号：2498583
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2498583.html
