基于学习的自动驾驶LIDAR定位

Learning-based LIDAR Localization for Autonomous Driving

论文提出DL-Net一种新颖的基于学习的LIDAR定位系统,创新地实现了各种深度神经网络结构的使用,建立基于学习的方法。DL-Net学习专门针对不同现实驾驶场景中的匹配而优化的局部描述符,并且高度了解全局上下文,这是深度学习的重要提示。在解决方案空间中建立的损失量上的3D卷积显着提高了定位精度,3D CNN通过将堆叠的沙漏网络(hourglass network)与中间监督(intermediate supervision)结合起来,去调整匹配损失量。RNN被证明可以有效地对车辆动力学进行建模,并具有更好的时间平滑性和准确性。为了证明该全球定位系统的性能和有效性,在KITTI数据集与其他算法进行了比较,并在长期的多重区段数据集上进行了评估。结果表明,论文的系统可以达到较高的精度。

In this paper,DL-Net,a novel learning-based LIDAR localization system,which innovatively implements the use of various deep neural network structures to establish learning-based methods. DL-Net learns local descriptors that are specifically optimized for matching in different real-world driving scenarios and has a high knowledge of the global context,which is an important cue for deep learning. The 3D convolution on the loss volume established in the solution space significantly improves the localization accuracy,and the 3D CNN adjusts the matching loss volume by combining a stacked hourglass network with intermediate supervision. RNNs are proven to effectively model vehicle dynamics with better temporal smoothness and accuracy. To demonstrate the performance and effectiveness of this GPS,it is compared with other algorithms on the KITTI dataset and evaluated on a long-term multi-segment dataset. The results show that the system of the paper can achieve high accuracy.