面向无人驾驶的多尺度语义地标同时定位与建图方法研究

The high precision positioning and the automatic construction of HD driving map in GNSS denied complex urban environment has been a hot and difficult problem in autonomous driving. It has been acknowledged by researchers that such a task should be realized through the vehicle's own perceptual means, based on the Simultaneous Localization and Mapping (SLAM) method. However, the traditional SLAM method is mainly oriented to robot application in homogeneous space, while ignoring human’s multi-scale perceptual and cognitive abstraction of landmarks in geographical space, which leads to perceptual or scale aliasing, thus could not guarantee its robustness in real driving scenes. To address the above problems, this application is intended to introduce the multi-scale abstraction and memory association of spatial knowledge into the SLAM method: the concept of Multi-scale semantic landmarks is proposed, and the key problems of scale-specific "object" landmark recognition, the scale invariant association of landmarks and the progressive graph optimization of multiscale landmarks are studied. This application proposes to construct Multi-scale learning samples automatically using the three-dimensional urban model by realistic rendering and enhanced by cycle generative adversarial network, and realizes the self-learning of scale-invariant description of landmarks by using the variational encoder-decoder, and realizes multi-scale progressive optimization by introducing the collapsing and splitting process into the graph model. Finally, a SLAM theoretical framework based on the multi-scale representation of semantic landmarks will be established and validated by using our "TiEV" autonomous driving platform.

在GNSS信号受遮蔽的复杂城区环境中的实时高精度定位和驾驶地图自动构建，一直是无人驾驶研究的热点与难点。需要借助车辆自身感知手段，基于SLAM方法实现。然而，传统SLAM方法主要面向同构的单一尺度空间中的机器人应用，忽视了地理空间中对地标的多尺度感知、认知的抽象过程，从而导致感知混淆和尺度混淆问题，无法保证其在真实驾驶场景中的鲁棒性。针对上述问题，本项目将人类对空间知识的多尺度抽象和关联记忆引入SLAM方法的研究中，提出多尺度语义地标概念，针对尺度特异的“对象”地标识别、地标的尺度不变关联以及多尺度地标的渐进式图优化关键问题分别展开研究。创新提出利用三维城市模型的真实感渲染和循环对抗生成网络自动构建多尺度学习样本，利用变分编码器实现尺度不变的地标描述自学习，将折叠和分裂过程引入图模型中实现多尺度渐进式优化。最终建立基于语义地标多尺度表达的SLAM理论框架，并在“途灵”无人车系统中进行验证。

高精度同时定位与建图（SLAM）是无人驾驶系统的关键技术基础。传统SLAM方法多依赖低层视觉特征或者激光点云的匹配，在大范围场景中低层特征易混淆，无法保证SLAM系统的鲁棒性。针对该问题，本项目从语义地标入手，探索利用语义地标的高层次抽象以实现高鲁棒性的无人驾驶语义SLAM系统。本项目探索了多种语义地标的表达，包括基于对象、基于图斑、基于人类可识别的语义特征以及基于点云聚类。并且创新提出了高性能的顾及多尺度特性的语义对象和语义特征的检测与分割网络，实现了各类语义地标的实时、准确识别。在此基础上，创新提出基于单目视觉的对偶二次曲面的语义地标的通用初始化方法和尺度估计方法，进而构建了基于对象语义地标的单目视觉SLAM系统。本项目创新提出了典型可识别语义地标的参数化方法和匹配方法，以及基于多假设的语义地标鲁棒关联方法。构建了基于典型可识别语义地标的实时SLAM系统，在无人驾驶平台上进行了部署验证。本项目面向大规模场景的视觉定位，提出了全局特异性图斑的识别方法以及基于图斑的加权匹配定位策略，实现了易混淆场景的全局定位。另外，还提出了基于地面语义特征的匹配和基于图神经网络的点云聚类空间关系不变性特征提取方法，实现了鲁棒的全局定位。在针对无人驾驶系统的SLAM研究中，本项目创新提出紧耦合车辆动力学模型的多传感器融合图优化方法。并将语义地标提供的约束与紧耦合SLAM系统结合，大幅提升了传统紧耦合SLAM方法在车载应用中的鲁棒性和精度。