Abstract: The proportion of target detection algorithms based on deep learning in the field of autonomous driving is increasing. This article first introduces the development process of convolutional neural networks and target detection algorithms based on deep learning. Briefly introduce the structural characteristics of several classic convolutional neural networks. A series of two-stage algorithms based on candidate frames represented by the R-CNN series and a regression-based one-stage algorithm represented by the YOLO series are introduced in detail. Briefly introduce the respective structures, advantages and disadvantages of these two types of algorithms. Finally, several commonly used optimization methods and research trends in the application of target detection algorithms in autonomous driving scenarios are summarized.

Key words: Deep learning; Target detection; Autonomous vehicles

摘要： 基于深度学习的目标检测算法在自动驾驶领域的比重日益上升。文章首先介绍了基于深度 学习的卷积神经网络和目标检测算法的发展过程，其中简要介绍了几种经典卷积神经网络模型的 结构特点；然后详细介绍了以 R-CNN 系列为代表的基于候选框的 two-stage 算法和以 YOLO 系列 为代表的基于回归的 one-stage 算法，简要介绍了这两大类算法各自的结构和优缺点，最后总结了 目标检测算法在自动驾驶场景中应用时比较常用的几种优化方法和研究趋势。

关键词: 深度学习；目标检测；自动驾驶