【作者】 李贤；

【导师】 李琳辉；

【作者基本信息】 大连理工大学 ， 车辆工程（专业学位）， 2019， 硕士

【摘要】 当今社会,能源短缺环境污染的大氛围、人工智能深度学习的大浪潮、汽车行业的大转型都促使着车辆的智能化发展。无人驾驶汽车是智能车辆的重要组成部分,行为决策是其三大核心技术之一。本文面向基于行人意图识别的智能车辆决策方法开展研究。首先,进行人体姿态估计的研究。对当前人体姿态估计的研究意义和方法进行简要分析,具体研究人体姿态估计的理论算法,包括:神经网络结构、关键点检测置信图、关键点连接部位亲和力场以及利用亲和力场进行多人解析,实现姿态估计在不同场景中的人体检测。其次,提出行人意图的检测方法。在人体姿态估计的基础上研究行人意图检测的可行性,采取切实可行的行人意图分类方法,提出大致的检测流程以供具体探究,对当前的评估方案进行分析。再次,完成智能车辆决策实验平台的搭建。搭建基于电动汽车的模拟实验平台,平台整体架构包含四个部分:环境感知模块、上位机决策模块、下位机控制模块、车辆执行模块,基于Ubuntu的上位机用于决策、下位机基于微控制器STM32F103ZET6进行开发,它们通过串口通讯,车辆速度控制部分以及制动部分的设计取决于微控制器的DAC和TIM外设。最后,完成车辆智能决策的模拟实验。制作行人意图数据集,以此作为智能行为决策的数据支撑,通过样本数据的训练和测试建立基于神经网络的决策模型,将上述过程的思想与车辆控制模块衔接,构成智能车辆决策系统。更多还原

【Abstract】 In today’s society,the atmosphere of energy shortage and environmental pollution,the wave of deep learning and artificial intelligence,and the great transformation of the automotive industry all promote the development of intelligent vehicles.Unmanned vehicle is an important part of intelligent vehicle,behavior decision-making is one of its three core technologies,This paper focuses on Intelligent Vehicle decision-making method based on pedestrian intention recognition.Firstly,human pose estimation is studied.The significance and methods of human pose estimation are briefly analyzed;the theoretical algorithm of human pose estimation is studied,Including: neural network structure、key point detection confidence map、part affinity field of key point connection、applying part affinity field to multi-person analysis;pose estimation for human detection in different scenes is realized.Secondly,a method for detecting pedestrian intentions is proposed.Feasibility of pedestrian intentions detection is studied based on human pose estimation;feasible behavioral intention classification method is adopted;a general detection process is proposed for specific exploration;and the current evaluation scheme is analyzed.Thirdly,the experimental platform of intelligent vehicle decision-making is built.Build a simulation experiment platform based on an electric vehicle;the overall framework of the platform consists of four parts: environment perception module,upper computer decision module,lower computer control module and vehicle execution module;upper computer based on Ubuntu is used for decision-making;lower computer is developed based on microcontroller STM32F103ZET6,which communicates through serial port;The design of vehicle speed control part and braking part depends on the DAC and TIM peripherals of the microcontroller.Finally,the simulation experiment of vehicle intelligent decision-making is completed.Make pedestrian intention dataset as data support for decision-making;the decision-making model based on neural network is established through training and testing of sample data;connect the above process ideas with vehicle control module to form an intelligent vehicle decision-making system.更多还原