LiDAR-based localization in a city-scale map is a fundamental question in autonomous driving research. As a reasonable localization scheme, the localization can be performed by global retrieval (that suggests potential candidates from the database) followed by geometric registration (that obtains an accurate relative pose). In this work, we develop a novel end-to-end, deep multi-task network that simultaneously performs global retrieval and geometric registration for LiDAR-based localization. Both retrieval and registration are formulated and solved as regression problems, and they can be deployed independently during inference time. We also design two mechanisms to enhance our multi-task regression network’s performance: residual connections for point clouds and a new loss function with learnable parameters. To alleviate the common phenomenon of vanishing gradients in neural networks, we employ residual connections to support constructing a deeper network effectively. At the same time, to solve the problem of huge differences in scale and units between different tasks, we propose a loss function that can automatically balance multi-tasks. Experiments on two public benchmarks validate the state-of-the-art performance of our algorithm in large-scale LiDAR-based localization.

城市规模地图中基于LiDAR的定位是自动驾驶研究中的一个基本问题。作为一种合理的定位方案，可以通过全局检索（从数据库中建议潜在的候选对象），然后进行几何配准（获得准确的相对姿态）来执行定位。在这项工作中，我们开发了一种新颖的端到端深层多任务网络，该网络可以同时为基于LiDAR的定位执行全局检索和几何配准。检索和注册都被公式化并解决为回归问题，并且它们可以在推理期间独立部署。我们还设计了两种机制来增强多任务回归网络的性能：点云的剩余连接和具有可学习参数的新损失函数。为了缓解神经网络中梯度消失的常见现象，我们采用残差连接来有效地构建更深的网络。同时，为了解决不同任务之间规模和单位差异巨大的问题，我们提出了一种损失函数，可以自动平衡多任务。在两个公共基准上进行的实验验证了我们的算法在基于LiDAR的大规模本地化中的最新性能。