Abstract We present a novel method to localize the vehicle on an easily accessible geo-referenced satellite image based on LiDAR. We first design a neural network to extract and compare the spatial-discriminative feature maps of the satellite image patch and the LiDAR points, and obtain the probability of correspondence. Then based on the outputs of the network, a particle filter is used to obtain the probability distribution of the vehicle pose. This method can use LiDAR points and any type of odometry as input to localize the vehicle. The experimental results show that our model can generalize well on several datasets. Compared with other methods, ours is more robust in some challenging scenarios such as the occluded or shadowed area on the satellite image.

中文翻译：

---

通过神经网络在卫星图像上进行基于激光雷达的车辆定位

摘要 我们提出了一种基于 LiDAR 在易于访问的地理参考卫星图像上定位车辆的新方法。我们首先设计了一个神经网络来提取和比较卫星图像块和激光雷达点的空间判别特征图，并获得对应的概率。然后基于网络的输出，使用粒子滤波器获得车辆姿态的概率分布。这种方法可以使用 LiDAR 点和任何类型的里程计作为输入来定位车辆。实验结果表明，我们的模型可以在多个数据集上很好地泛化。与其他方法相比，我们的方法在一些具有挑战性的场景中更加鲁棒，例如卫星图像上的遮挡或阴影区域。