【作者】 李锦峰；

【导师】 裴伟；

【作者基本信息】 大连海事大学 ， 工程硕士（专业学位）， 2023， 硕士

【摘要】 近年来,随着汽车数量的增加,无人驾驶技术的发展变得越发紧迫。无人驾驶技术是指汽车利用自动驾驶系统感知环境、自主规划路径、避开障碍物,从而实现自主航行的过程。双目深度估计和车辆转向预测分别是自动驾驶系统感知模块和决策模块中的关键任务。然而,现有基于深度学习的双目深度估计和车辆转向预测算法存在计算量大、功耗高、自适应能力差等问题。相比之下,脉冲神经网络作为第三代神经网络,具有时间连续性、计算能耗低、快速处理、仿生性强等优点,与自动驾驶系统中双目深度估计和车辆转向预测任务的需求高度契合。因此,本文将基于脉冲神经网络,在现有研究工作的基础上,重点研究自动驾驶系统中的双目深度估计和车辆转向预测任务。本文的主要工作如下:(1)针对基于脉冲神经网络的Stereo＿Spike算法输出深度图存在的不适定区域边缘模糊、场景轮廓不明显以及局部细节丢失等问题,本文提出了具有语义约束的相关联层、参数融合的逐元素脉冲残差模块、基于脉冲形式的空洞空间池化金字塔模块以及边缘损失四点改进方案。实验结果表明,改进后的模型在能耗、深度估计准确性以及实时性方面均具有较好的表现。(2)针对现有车辆转向预测模型能耗高和自适应能力差的问题,本文提出了一种基于脉冲神经网络的车辆转向预测模型。该模型使用参数漏积分点火模型作为网络中的脉冲神经元,以端对端的方式将事件相机输出的事件流映射到车辆转向角,并通过梯度替代法实现整个模型的参数更新。事件相机与脉冲神经网络的结合,使得该模型能够像人类驾驶员一样更加关注场景中的动态特征,从而提高其在不同环境和情况下的自适应能力。实验结果表明,该模型在能耗、自适应能力以及预测准确性方面均有较好的表现。(3)本文使用LGSVL作为基础工具,设计并实现了一套基于人在回路的无人驾驶仿真系统。在研究内容(1)和研究内容(2)的基础上,构建了该系统中的自动驾驶系统模块,并通过基于人在回路的仿真测试模块进一步提升车辆转向预测模型在不同环境和情况下的自适应能力。同时,通过该系统进一步验证本文所提方法的有效性。更多还原

【Abstract】 In recent years,with the increase in the number of cars,the development of driverless technology has become more urgent.Unmanned driving technology refers to the process in which a car uses an automatic driving system to sense the environment,plan its own path,and avoid obstacles,thereby realizing autonomous navigation.Binocular depth estimation and vehicle steering prediction are key tasks in the perception module and decision module of the autonomous driving system,respectively.However,the existing binocular depth estimation and vehicle steering prediction algorithms based on deep learning have problems such as large amount of calculation,high power consumption,and poor adaptive ability.In contrast,as the third-generation neural network,the spiking neural network has the advantages of time continuity,low computing energy consumption,fast processing,and strong bionicity.It is compatible with the needs of binocular depth estimation and vehicle steering prediction tasks in automatic driving systems.Highly fit.Therefore,this thesis will focus on binocular depth estimation and vehicle steering prediction tasks in autonomous driving systems based on spiking neural networks and building on existing research work.The main work of this thesis is as follows:(1)Aiming at the problems of blurred edges in ill-posed regions,inconspicuous scene outlines,and loss of local details in the depth map output by the Stereo＿Spike algorithm based on spiking neural networks,this thesis proposes an associative layer with semantic constraints and an element-by-element spiking residual module with parameter fusion,Pyramid module of empty space pooling based on pulse form and four-point improvement scheme of edge loss.Experimental results show that the improved model has better performance in terms of energy consumption,depth estimation accuracy and real-time performance.(2)Aiming at the problems of high energy consumption and poor adaptive ability of existing vehicle steering prediction models,a vehicle steering prediction model based on spiking neural network is proposed in this thesis.The model uses the parameterized leaky integral ignition model as the spiking neuron in the network to map the event stream output by the event camera to the vehicle steering angle in an end-to-end manner,and realizes the parameter update of the entire model through the gradient substitution method.The combination of the event camera and the spiking neural network enables the model to pay more attention to the dynamic features in the scene like a human driver,thus improving its adaptive ability in different environments and situations.The experimental results show that the model has a good performance in terms of energy consumption,adaptive ability and prediction accuracy.(3)This thesis uses LGSVL as the basic tool to design and implement a set of unmanned driving simulation system based on human-in-the-loop.On the basis of research content(1)and research content(2),the automatic driving system module in the system is constructed,and the simulation test module based on human-in-the-loop is used to further improve the performance of the vehicle steering prediction model in different environments and situations.Adaptability.At the same time,the effectiveness of the proposed method is further verified by the system.更多还原