标题： 合成与识别自动驾驶汽车摄像雷达数据集中的噪声水平 显示英文标题

标题： Synthesizing and Identifying Noise Levels in Autonomous Vehicle Camera Radar Datasets

摘要： 检测和跟踪对象是任何自主导航方法的关键组成部分。在过去的几十年里，使用神经网络在各种数据集上已经取得了有希望的检测结果。虽然许多方法侧重于性能指标，但很少有项目致力于提高这些检测和跟踪管道的鲁棒性，尤其是在传感器故障方面。在本文中，我们试图通过为相机-雷达自动驾驶汽车（AV）数据集创建一个现实的合成数据增强管道来解决这个问题。我们的目标是准确模拟由于现实世界干扰导致的传感器故障和数据退化。我们还展示了在我们增强的数据集上训练和测试的轻量级噪声识别神经网络的结果，达到了在10086张图像和2145个雷达点云上的11个类别中总体识别准确率为54.4\%。 显示英文摘要

摘要： Detecting and tracking objects is a crucial component of any autonomous navigation method. For the past decades, object detection has yielded promising results using neural networks on various datasets. While many methods focus on performance metrics, few projects focus on improving the robustness of these detection and tracking pipelines, notably to sensor failures. In this paper we attempt to address this issue by creating a realistic synthetic data augmentation pipeline for camera-radar Autonomous Vehicle (AV) datasets. Our goal is to accurately simulate sensor failures and data deterioration due to real-world interferences. We also present our results of a baseline lightweight Noise Recognition neural network trained and tested on our augmented dataset, reaching an overall recognition accuracy of 54.4\% on 11 categories across 10086 images and 2145 radar point-clouds.