【作者】 陈卓；

【导师】 路松峰；

【作者基本信息】 华中科技大学 ， 计算机技术， 2019， 硕士

【摘要】 作为未来汽车工业的发展方向,智能驾驶是必然趋势。汽车的驾驶与人工智能的结合越来越紧密,在感知层面对环境进行分类、目标检测、语义分割等是结合的主要运用,在算法运用部分从机器学习到深度学习发生了巨大的变迁。从传统的HOG和SVM、SIFT等到深度学习的卷积神经网络、循环神经网络、递归循环网络等,因为深度学习带来可见的效果,被广泛运用在自动驾驶的环境感知层面。基于深度学习神经网络的基础上通过将交通拥堵时的车辆低速运动状态抽象为跟驰模型(Car Following Model),在树莓派硬件基础上,通过一系列的数据采集,网络模型搭建,实体模型搭建,训练调优,设计和构建自动跟驰驾驶框架,为车辆在单车道拥堵的路况上提供了一种有条件的自动驾驶思路。主要实现车辆的自适应巡航、对交通信号的识别处理,对车辆的识别以及自适应跟驰三个功能。运用端到端的感知到控制策略,设计搭建了卷积神经网络,通过记录操作者的序列,训练网络模型,直接输出对车辆的控制,作为一级控制策略,实现了车辆的自适应巡航。在DarknetYolov3目标检测网络的基础上模拟车辆在城市交通运行的环境,通过对模型的Finetune重新训练,设计实现了二级控制策略,模拟了车辆在城市单车道交通环境下的自动跟驰驾驶。更多还原

【Abstract】 As the future development direction of automobile industry,intelligent driving is an inevitable trend.The combination of automobile driving and artificial intelligence is getting closer and closer.The main application of the combination is the classification of environment,target detection and semantic segmentation at the level of perception.In the part of algorithm application,great changes have taken place from machine learning to deep learning.HOG,SVM,SIFT and other deep learning convolutional neural network,circular neural network and recursive circular network are widely used in the environmental perception level of autonomous driving because of the visible effect brought by deep learning.Based on the deep study on the basis of neural network through the traffic jam of the vehicle at low speed when the motion state abstraction for the Following Model(Car Following Model),on the basis of raspberries pie hardware,through a series of data collection,network Model,entity Model building,training tuning,design and construction of automatic Following driving framework,is for the vehicle in the lane congestion on the roads of provides a conditional autopilot.The main functions are adaptive cruise,traffic signal recognition and processing,vehicle recognition and adaptive galloping.Using the end-to-end sensing control strategy,the convolutional neural network is designed and constructed.By recording the operator’s sequence,training the network model,and directly outputting the control of the vehicle,as a first-level control strategy,the adaptive cruise of the vehicle is realized.On the basis of darknet-yolov3 target detection network,the vehicle operating environment in urban traffic is simulated.Through the Finetune training of the model,the two-level control strategy is designed and realized,and the vehicle’s automatic driving in the urban single-lane traffic environment is simulated.更多还原