城市场景动态激光点云的时域关联随机场分析

The scans of 3D Velodyne LiDAR are 360 panoramic views, overlapped and correlated, which can be constituted into long sequence. Aimed at those characteristics, the temporal informations of point cloud are introduced into the Label-sets of nodes in Markov Network. Meanwhile, the coherences of motion/attribute/appearance of dynamic point cloud sequence will be imported into process of Markov Network optimization with time domain association and the sliding window. .Therefore, the recognition method of point-by-point classification and coherence tracking is realized in this project approximately. It became the novel associated Markov network point cloud classification model with capability of temporal dynamic information analysis..Innovations are: .1) This topic of time domain association research has not been reported in existed papers of point cloud classification. In addition, the model can achieve a better point by point the precise classification and movement analysis result in contrast with other moving targets tracking methods;.2) In the thought of this research, the time domain related information of dynamic point cloud in the driving environment is used both in node’s level and sequence’s level in this model. It can be greatly improved the accuracy and effectiveness of classification result of Markov network both in spatial and in temporal consistence..Eventually, the algorithm will be used for mobile mapping system, which can both analysis static and dynamic targets in urban driving environment automatically and improve accuracy and speed of the professional data extraction process greatly.

本项目针对无人驾驶车辆的Velodyne雷达采集动态点云的特点(如：360度全景、高重叠度、前后帧关联、动态长时间序列)，引入时域关联信息到马尔科夫随机场点云分类模型的节点类别标记集中；并将点云帧间类别、形态变化的一致性代入到下一个滑动窗口随机场的优化过程中。近似实现了一个逐点同时进行点云分类和类别变化跟踪的动态分析识别方式，成为一种新颖的具有时域动态信息分析能力的关联马尔科夫网点云分类分析模型。.创新体现在：1）在选题上，选择现有点云分类研究并没有涉及的时域关联方向，而且对比已有点云动目标跟踪方法来说，该模型可实现更精细的逐点精确分类和运动分析；2）在思路上，在节点和序列两个不同层次上，基于关联马尔科夫网模型引入、分析、表达时域关联信息，极有潜力在时空一致性上大幅度提高动态点云分类的精确度和有效性。.算法最终将用于移动测图系统的自动数据加工，为无人驾驶提供高精度全息的地图数据。

本项目针对无人驾驶车辆的Velodyne雷达采集动态点云的特点(如：360度全景、高重叠度、前后帧关联、动态长时间序列)，引入时域关联信息到马尔科夫随机场点云分类模型的节点类别标记集中；并将点云帧间类别、形态变化的一致性代入到下一个滑动窗口随机场的优化过程中。近似实现了一个逐点同时进行点云分类和类别变化跟踪的动态分析识别方式，成为一种新颖的具有时域动态信息分析能力的关联马尔科夫网点云分类分析模型。.基于以上的算法创新，并结合深度学习的迅猛发展，项目完成了真实场景VeloSLAM+DATMO，实现将点云帧间类别、形态变化的一致性代入到下一个滑动窗口随机场的优化过程中，参与场景分割和点云匹配过程，相关代码和效果已经公开。随后本研究实现动态点云的3D深度学习，实现图像和点云特征融合的3D目标分割，相关算法在KITTI性能对比优异，已经用于东风和上汽的自动驾驶车辆，用于为无人驾驶提供高精度全息的地图数据，同时简化算法用于实时的路面范围探测和动态目标分割。