面向无人驾驶的多源感知数据概率面模型表达与语义理解

Multi-source sensors based road scene perception refers to use perception sensors, localization sensors and map sensor to detect the moving/static objects, drivable areas, position information and traffic state, accurately and steadily, which is one of the core technologies of autonomous driving, could determine the intelligent level of the unmanned vehicle, directly. How to fuse monocular camera, stereo camera, Lidar, Radar and high-precision map for three-dimensional space expression, and use this model for fast and accurate space-time level semantic understanding of the driving environment is a difficult problem. By extending data fusion theory and robotics navigation and mapping, this project proposes a heterogeneous spatial-temporal perception data for the study of probabilistic plane model for the universal expression of multi-source heterogeneous 3D perception data. We propose a 3D mapping method based on plane-based clustering and deep convolution neural network. Then, we propose to use the combination of the depth of the network and the conditional random field to extend the semantic segmentation of the road from the single dimension of a single time to the multi-dimensional semantic understanding. This study can break new ground in fusion theory and perception method in the field of autonomous driving, lay a new theoretical foundation and provide technical support for unmanned vehicle, high-precision maps, robot navigation, etc.

基于多源传感器融合的道路场景实时精准感知是指通过车载感知定位或地图传感器实现对当前驾驶场景的动静态物体、可行驶区域、位置信息以及交通规则状态信息的准确实时的理解，是无人驾驶最为核心技术之一，直接决定无人驾驶车的智能程度。如何充分融合单双目相机、激光雷达、毫米波雷达以及高精度地图等多源异构感知数据的属性优势，形成更合理的三维空间表达模型，同时基于此模型进行快速精准的驾驶环境的深度动态语义理解是一个难点。本课题以多源、异构的驾驶场景感知数据融合作为研究切入点，拓展数据融合理论和机器人导航建图理论，研究多源异构三维感知数据一体化的概率面模型表达方法；提出基于非监督学习的面聚类和深度卷积神经网络的三维微重建方法；通过深度对抗网络和条件随机场相结合的方法把道路语义分割从单一时刻单一维度扩展到包含实序信息的多维度语义理解，为无人驾驶、高精度地图、机器人导航等提供新的理论基础和技术支撑。

基于多源传感器融合的道路场景感知是无人驾驶最为核心技术之一，直接决定无人驾驶车的智能程度。但是多源传感器的数据格式不同，采样频率不同；如何充分融合多源异构感知数据的优势，进行快速精准的动态语义理解是该技术一个难点。本项目探索了多传感器数据的时间对齐方法，建立了多源异构数据一体化模型；提出了异构数据的三维点云重建方法，实现了高精度的单智能体建图和多智能体协同建图；研究了道路场景多维度理解方法，实现了复杂道路场景下高精度的车辆检测、跟踪和语义分割；构建了多个大型数据集，并结合实际场景对提出的方法进行了实验。通过项目的开展，发表相关论文28篇，申请发明专利7项，达到了预期研究目标。