顾及扫描上下文的预测与判决相结合的点云在线分类方法

3D laser scanning technique is being widely used for scene understanding and object classification. Online classification of point cloud data could be used for route planning and barrier avoiding of unmanned platforms based on semantic information it provided. Due to abundant features used and fixed structures of classifiers which are not changed based on scene changing, the efficiency and precision of classification could not be both satisfied by using traditional classification methods of point cloud data. Based on scanning context, this project investigates an online updating classifier which combining prediction and decision along with the objects' classes changing in scene. First, a class transmission model is designed based on spatial relationship of objects. The classes of objects on route are predicted using the classified objects which have been scanned previously. Second, a non-balanced decision tree is constructed for efficient classification according to the classes probability of prediction. This could be improve the classification efficiency due to reducing the amount of irrelevant features. Finally, in order to demonstrate the applicability of proposed models and algorithms, a prototype online classifciation system is developed and many tests are designed for analyzing the classification precision and efficiency. The development of fine modelling will provides a new method for high-precision classification efficiently. It is expected that methods developed in this project not only provide technique for scene understanding and navigation of unmanned platforms, but also promote the theory level of classification of point cloud data.

三维激光扫描技术目前广泛应用于场景理解及地物类别区分，在线点云分类能够为无人驾驶及导航提供行进路线上场景语义信息，用于路线规划及避障。传统点云分类方法，由于特征使用量多，在分类过程中未依据场景变化在线更新分类结构，不能同时满足在线分类效率及精度要求。本课题结合扫描上下文地物联系，采用先预测后判决的思想，研究随场景中地物类别变化，在线更新的分类优化方法及模型。为此需要研究：构建基于空间相关性的类别传播模型，依据扫描顺序中已分类地物，对行进路线上的待分类地物的类别概率进行预测；构建类别非均衡的决策树分类模型，按照待分类对象预测类别的概率，依次从高到低对可能的类别进行序列判决，减少分类过程中非相关类别特征使用量。最后，搭建实验平台，对模型和算法进行综合试验验证。本课题能够解决兼顾分类精度及效率的点云在线分类问题，为场景感知及无人导航提供技术支撑，提升基于点云场景理解的理论水平。

本项目以地图与导航服务的实际需求为导向，研究了多传感器场景采集及在线数据生成、基于压缩感知的地物特征提取及分类技术、基于上下文信息提取、序列化在线分类技术等多方面的关键技术，形成了一整套从数据获取、理论方法到实用算法的点云分类及地物识别的方案。本项目的成果能够应用到与导航与位置服务等领域相关的各种应用场景，并为相关领域的研究提供基础理论与方法支撑。.首先搭建实验平台。本项目重点研究了设备小型化的基础上，选取合适的传感器，集成多传感器（包括激光与图像采集器、定位定姿传感器等）用来获取地物的完备性信息。研制高精度时间同步芯片，记录各种传感器在信号获取时刻的时间，并且将这些时间进行同步处理，为多信号融合提供基础。研究了在地物上下文信息及地物结构约束下的，激光辅助惯导的SLAM算法，能够精确获取机器人平台的运行轨迹，为后续的地图生成和模型建立提供基础。.针对特征使用量多，本项目研究了基于压缩感知的地物特征提取及分类方面：（1）研究使用非均衡多尺度分割的方法描述多层次地物特征，近而能够获取高效率、高精度的地物分类结果。（2）结合词袋和压缩感知以及稀疏表达的方法对特征进行压缩。同时，引入压缩感知以及稀疏表达的思想，用于特征压缩。通过构建字典，每一点特征就可以使用字典中词语的出现频次表示。这样的表示能够改善原始数据冗余以及不规则的缺陷，起到特征选择以及压缩的作用。.针对在线分类模型，本项目研究了序列化在线分类技术，结合上下文地物关系的在线分类方法，构建序列化分类器，并且将多层次地物分类结果进行综合，获取高精度分类结果。序列化分类器自动特征选择、适合处理多类别分类中的大量的特征输入。.本课题能够解决兼顾分类精度及效率的点云在线分类问题，为场景感知及无人导航提供技术支撑，提升基于点云场景理解的理论水平。