【作者】 周林；

【导师】 刘国金；

【作者基本信息】 重庆大学 ， 信息与通信工程， 2020， 硕士

【摘要】 随着当今世界科学技术的进步,人工智能也在飞速地发展。无人驾驶技术自20世纪70年代提出以来,一直是人工智能发展的一个重要领域。无人驾驶技术主要包含感知、决策和执行三个部分。而感知则是整个无人驾驶技术的前提和基础,只有在实现了感知的基础上,才能做出相应的路径规划和驾驶行为决策。顾名思义,感知就是感受、观察车辆周边的路况环境并对其进行理解认知,能否全面有效的对环境进行感知直接影响了整个无人驾驶系统的工作性能。无人驾驶车辆常用的感知传感器有视觉传感器(单双目摄像头、深度相机等)、毫米波雷达、激光雷达或是由这些传感器进行不同组合形成的感知系统等。现阶段量产应用的无人驾驶系统多采用基于单传感器的环境感知算法,而由于传感器器件各有利弊,使用单传感器进行环境感知存在明显的局限性。通过多传感器信息冗余和优势互补,使用多传感器搭配环境感知系统,可以明显解决现有的问题。因此,本文研究了毫米波雷达和摄像头的多源信息融合,基于融合思想设计了一个环境感知算法,包括数据解析、目标初选、目标检测和跟踪、传感器采样周期校准和目标数据融合等。主要内容如下:(1)对毫米波雷达和摄像头原始数据进行了控制器局域网总线(CAN,Controller Area Network)协议解析,在目标初选中,对原始数据设计了目标筛选算法,能够有效滤除毫米波雷达原始数据中的空目标、杂波噪声信号和虚假虚警目标。在目标检测中,目前常用的算法是卡尔曼滤波。卡尔曼滤波算法适用于线性、离散和有限维空间,它能够从包含噪声的观察序列中预测出物体的坐标位置及速度,被广泛运用于雷达、计算机视觉等一些工程应用中。但是,经典的卡尔曼滤波算法对目标检测和在实际应用中都具有局限性,本文提出了带一致性检验和生命周期决策的卡尔曼滤波,可以从原始传感器信号中提取出有效的危险目标,并提升了目标检测和跟踪的效果。(2)针对毫米波雷达和视觉摄像头信号采集周期不同的问题,本文分析了这两种传感器在实际应用中的工作特性,建立了运动学方程并采用了改进的内插外推法对传感器的感知矢量进行校准和补偿,完成了异源传感器在时间上的同步,为传感器数据进行进一步融合奠定了基础,提升了感知系统的检测能力。(3)在多传感器信号融合中,目前常用的算法有加权法、D-S推论、贝叶斯估计、航迹关联算法等。本文从异构传感器数据融合的可行性出发,采用了状态矢量融合算法对毫米波雷达和摄像头传感器的数据进行信号融合,使得整个感知系统整合了毫米波雷达和摄像头传感器各自的优势,增强了整个环境感知算法的性能。(4)本文搭建了Matlab/Simulink和dSPACE的联合仿真平台对在中国汽车工程研究院采集的道路数据集进行了仿真实验。仿真结果表明,本文提出的基于毫米波雷达和摄像头传感器的多源信息融合环境感知算法能够有效感知自动驾驶汽车外部的路况环境,其实时检测能力能得到充分的保证。同时经实车试验后的可视化结果表明,验证了该算法在无人驾驶实际应用中的有效性和稳定性。更多还原

【Abstract】 With the progress of science and technology in today’s world,artificial intelligence is also developing rapidly.Since it was put forward in the 1970 s,autonomous driving technology has been an important field in the development of artificial intelligence.The technology of autonomous driving mainly includes three parts: perception,decision-making and execution.Perception is the premise and foundation of the whole autonomous driving technology,Only on the basis of realization of perception can we make corresponding path planning and driving behavior decision.Just as the name implies,perception is to feel and observe the road conditions surrounding the automated vehicle and recognize them.The common perception of sensors of automated vehicles are visual sensors(monocular and binocular cameras),millimeter wave radar,lidar and the perception system formed by different combinations of these sensors,etc.At present,most autonomous driving system in mass production use the environment perception algorithm based on single sensor.However,due to the advantages and disadvantages of each sensor,there are obvious limitations in using single sensor for environment perception.By using multi-sensor information redundancy and complementary advantages,the multi-sensor environment perception system can obviously solve the existing problems.Therefore,this paper studies the multi-source information fusion of millimeter wave radar and camera,and designs an environment perception algorithm based on the data fusion,including data analysis,target primary selection,target detection and tracking,sensor sampling cycle calibration and target data fusion.The main contents are as follows:(1)Using CAN protocol to analyze the original data of Millimeter wave radar and camera.In the target primary selection,this paper designs the target filtering algorithm,which can effectively filter out the empty target,noise signal and false alarm object in the original data of MMW radar.In the target detection,the commonly used algorithm is kalman filter.Kalman filter algorithm is suitable for linear,discrete and finite dimensional space,it can predict the coordinate position and velocity of objects from the observation sequences containing noise,and it is widely used in radar,computer vision and other engineering applications.However,the classical kalman filter algorithm has limitations on target detection and in practical application.This paper proposes a kalman filter with consistency check and life cycle decision-making,which can extract effective dangerous target from the original data and improve the accuracy of target detection and tracking.(2)Aiming at the problem of different signal sampling cycles of millimeter-wave radar and visual camera,we analyze the operating characteristics of these two sensors in practical applications,then set up kinematic equation and use an improved interpolation extrapolation method to calibrate and compensate the sensors’ perception vector,complete the time synchronization of heterogeneous sensors,which laid the foundation for further data fusion and improve the detection ability of perception system.(3)In multi-sensor signal fusion,the commonly used methods at present include weighting method,D-S inference,Bayesian estimation and track correlation algorithm,etc.In this paper,we start from the feasibility of heterogeneous sensor data fusion,adopt the decision-making level state vector fusion algorithm to fuse the signal data of millimeter wave radar and camera sensor,which make the perception system integrates the advantages of millimeter wave radar and camera sensor,enhance the performance of the whole environment perception algorithm.(4)In this paper,we build a joint simulation platform of Matlab/Simulink and dSPACE to simulate the road dataset obtained in China Automotive Engineering Research Institute Co.,Ltd.The simulation result shows that our proposed multi-source information fusion environment perception algorithm based on millimeter wave radar and camera sensor in this paper can effectively perceive the road environment outside the autonomous vehicle,and the real-time detection ability can be guaranteed also.At the same time,the visualization results after the real vehicle test indicates that the effectiveness and stability of the algorithm in the practical application of automated vehicle are verified.更多还原