【作者】 李白云；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2011， 博士

【摘要】 机器人在国防、民用、科研等领域都有着广泛的应用前景,在计算机技术里一直处于研究的前沿。机器人的自主导航系统庞大、技术复杂,在研究和实验过程中迫切需要真实度高、实时性好的仿真系统以验证各类算法的有效性,为智能感知,路径规划等模块提供预研与技术验证等支持。本文针对构建面向机器人自主导航技术所需要的关键技术展开,以计算机图形学和虚拟现实技术为基础,进行了如越野环境的渲染,传感器仿真算法,以及虚拟实验环境的构造等研究,并取得一定成果,其中包括：针对野外越野场景的地形渲染的需要,提出了利用GPU实现的四叉树地形渲染算法,通过GPU提高四叉树的构建速度,在此基础上提出了以GPU实现的四叉树遍历,场景分割,裂缝消除和多边形化等技术。取得了较传统算法更为精确的渲染效果和3-4倍渲染速度的提升。同时,通过将处理大尺寸场景常用的clipmap技术引入四叉树构造技术中,拓展了该技术在大尺寸场景上的应用空间。并提出了一系列裂缝消除,数据调度和渲染优化策略。实验证明该算法结合了动态算法和固定网格算法的优点,没有预处理的需求,因此运行期对数据传输的依赖很低,同时几乎没有带宽需求,在低CPU负载的情况下可以获得很高的帧速率,是应用前景很广泛的地形渲染技术。针对传感器仿真的需要,本文提出利用GPU硬件实现的激光雷达和光学摄像机仿真技术。通过引入环境纹理、bump map和位移纹理等技术,实现这两类自主机器人上普遍使用的传感器的仿真。对激光雷达的实验表明,该算法普适性好,针对各类雷达都能很好地仿真,同时拥有满意的仿真速度。对光学摄像机的实验表明,仿真算法可以很好地建立虚拟摄像机和仿真摄像机之间的关系,真实度很好,同时算法还可以很好地模拟各类镜头的畸变。由于算法充分利用了GPU结构,本文的算法不仅在速度上有明显优势、可运行于普通桌面系统而无须复杂的大型系统,更重要的是这两种算法充分考虑了现代渲染技术对GPU的依赖,通过GPU执行的仿真算法不仅有效地节约了系统的带宽需要,同时也降低了CPU的负担。针对虚拟环境搭建的需要,本文对利用自主机器人上所携带的传感器进行虚拟场景搭建的技术进行了研究。其一是通过激光雷达实现对多边形场景的构建,通过对多边形的筛选,获得可靠的网格。并在此基础上执行网格的简化算法,构造可以满足渲染器渲染需要的网格。实验证明网格的简化算法可以有效地保留模型的特征,大幅度减少冗余多边形,停机策略良好。其二是通过激光雷达和光学传感器所生成的体视数据进行的越野环境数字高程数据的构建,以及平滑、填充等一系列操作,达到构建越野环境的目的,实验证明该算法简单有效,对空白区数据的填充效果良好。更多还原

【Abstract】 Autonomous robotic technology is wildly used in defense, civil economy and academic researches. As a research combined with image processing, data collection, data fusion, intelligent sensor, pattern recognition, automatic control and system simulation technologies, autonomous robotic technology is one of the hottest areas in computer science. Virtual simulation system as one strong support for intelligent sensor and path land modular is a key technology of autonomous robotic researches.This thesis is focus on technologies which is needed for construct the autonomous robot virtual simulation system. Our researches is based on computer graphic and virtual reality technology to development researches on terrain rendering, sensor simulation, and the construction of virtual experiment environments. The results include:To solve the outdoor rendering problems, we proposed a GPU-implemented quad-tree based terrain rendering scheme, increasing the speed of quad-tree construction by exploit the stream-structure of GPU. Base on the scheme, we propose GPU-implemented quad-tree traversing, scene division, crack fixing and triangulation schemes, to rendering the DEM, with good rendering results and frame rates. Moreover, by introducing the clipmap technology which is widely used in large scale scene rendering into quad-tree construction, we proposed a set of technology including crack fix strategy, data management scheme and rendering optimization strategy to solve the terrain rendering of large scale scene. The experiments prove that our scheme combined with the advantages of dynamic construction algorithm and fix mesh algorithm, need no preprocessing, lightly reliance to runtime data transmission, with little bandwidth consumption. The experiments also show that our scheme could reach high frame rates with low CPU payload.To satisfy the need of sensor simulation, this thesis has also proposed technologies for laser radar and optical camera simulation, which are the most popular sensors on autonomous robot, by exploiting environment texture and bump map features of GPU. The experiment of laser radar simulation shows that our algorithm is efficient for most radar type, with a satisfactory simulation speed, while the experiment of optical camera show that the virtual camera could build relationship with real camera with high reality, also could it simulate various lens distortions. As our method fully consider about the GPU architechture, it could not only reach very high frame rates in ordinary desktop platform, but also reduce the usage of bandwidth and payload of CPU. This thesis also presented our researches on virtual experiment environment construction with sensor on autonomous robot. One is construction of polygon environment with laser radar. With a scheme of polygon linkage, mesh simplification to construct the polygon meshes the experiments needed. Experiments show that our mesh simplification strategy could largely reduce the redundant polygons while maintaining the geometry feature of existing meshes, and with good halting strategies. The second is technology of constructing outdoor environment’s DEM data by data fusion with laser radar and height field data from optical cameras and a set of operation such as smoothing and blank area filling. Experiment show that our method was simple and efficient, and good at blank area filling.更多还原