Cooperation of automated vehicles (AVs) can improve safety, efficiency and comfort in traffic. Digital twins of Cooperative Intelligent Transport Systems (C-ITS) play an important role in monitoring, managing and improving traffic. Computing a live digital twin of traffic requires as input live perception data of preferably multiple connected entities such as automated vehicles (AVs). One such type of perception data are evidential occupancy grid maps (OGMs). The computation of a digital twin involves their spatiotemporal alignment and fusion. In this work, we focus on the spatial alignment, also known as registration, and fusion of evidential occupancy grid maps of multiple automated vehicles. While there exists extensive research on the synchronization and fusion of object-based environment representations, the registration and fusion of OGMs originating from multiple connected vehicles has not been investigated much. We propose a methodology that involves training a deep neural network (DNN) to predict a fused evidential OGM from two OGMs computed by different AVs. The output includes an estimate of the first- and second-order uncertainty. We demonstrate that the DNN trained with synthetic data only outperforms a baseline approach based on coordinate transformation and combination rules also on real-world data. Experimental results on synthetic data show that our approach is able to compensate for spatial misalignments of up to 5 meters and 20 degrees.

翻译：自动驾驶车辆（AVs）的协作能够提升交通安全、效率与舒适性。协同式智能交通系统（C-ITS）的数字孪生在交通监测、管理与优化中扮演关键角色。计算交通实时数字孪生需要以多个互联实体（如自动驾驶车辆）的实时感知数据作为输入，其中一类典型感知数据是证据占据网格图（OGMs）。数字孪生的计算涉及其时空对齐与融合。本文聚焦多辆自动驾驶车辆证据占据网格图的空间对齐（即配准）与融合问题。尽管针对基于对象的环境表征的同步与融合已有广泛研究，但对多辆网联车辆生成的OGMs的配准与融合研究仍较为匮乏。我们提出一种方法，通过训练深度神经网络（DNN），从两辆不同自动驾驶车辆计算的OGMs中预测融合后的证据占据网格图，其输出包含一阶与二阶不确定性的估计。实验证明，仅使用合成数据训练的DNN在真实数据上也能超越基于坐标变换与组合规则的基线方法。合成数据实验结果表明，我们的方法能够补偿高达5米和20度的空间错位。