【作者】 徐兵；

【导师】 梁军；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2020， 硕士

【摘要】 随着社会和经济的发展,无人驾驶的概念已逐渐走入日常生活。相对于有人驾驶系统,无人驾驶能够避免因驾驶员注意力不集中等原因导致的交通事故,并且能够解放驾驶员的双手。人们对于无人驾驶所带来的便捷性与安全性充满了向往,同样无人驾驶也吸引着广大研究人员的关注。但目前无人驾驶的实际落地仍然存在不小的挑战,其中亟待解决的主要问题包括:环境感知中障碍目标检测的精度,以及车辆自主决策的准确性与安全性。本文针对无人驾驶环境感知技术中存在的车辆障碍物检测精度低及速度慢等问题行了针对性的改进,并对车辆自由换道决策行为进行了建模分析。针对基于激光雷达的传统栅格化车辆目标检测方法中存在的聚类准确率低、超参数多等问题,在现有的快速密度峰值聚类算法(DPC)的基础上提出了基于相互近邻的自适应合并密度峰值聚类算法(MNN-ADPC);针对基于激光雷达的神经网络车辆目标检测方法中存在的点云特征提取不足以及损失约束不合理等问题,提出了优化特征提取模块与损失项的改进三维目标检测网络;针对车辆自由换道行为由于受影响因素较多而难以准确预测的问题,引入梯度提升决策树进行特征变换,结合逻辑回归,构建融合决策模型对车辆自由换道行为进行分析。本文的主要研究工作和成果如下:1)针对基于激光雷达的传统栅格化车辆目标检测方法中,障碍栅格聚类时采用的传统聚类方法存在聚类精度低及速度慢等问题,同时在实际应用中,超参数过多等问题严重影响了聚类算法准确性和适应能力。通过对现有的DPC算法的初始聚类中心选择、待分配点划分等策略进行改进,提出了基于相互近邻的自适应合并密度峰值聚类算法(MNN-ADPC)。改进的聚类算法采用较宽松的初始聚类中心选择方式,并结合后续的自适应合并策略,解决了采用固定阈值筛选方式存在的类簇中心漏选的问题。同时,针对原始快速密度峰值聚类算法中采用的最近邻划分策略存在的误划分情况,改进聚类算法中采用了更加鲁棒的相互近邻划分策略,解决了嵌入型数据中存在的误划分问题。通过在多个数据集上的对比实验,结果表明了所提出的基于相互近邻的自适应合并密度峰值聚类算法具有更好的准确性与适应性。同时,相较于原始的快速密度峰值聚类算法,改进的聚类算法具有更少的超参数量,无需人工参与。在真实的无人驾驶环境中,基于相互近邻的自适应合并密度峰值聚类算法能够获得更加准确的聚类效果。2)针对现有基于激光雷达的神经网络车辆目标检测方法中,三维车辆目标检测网络由于受到点云稀疏特性的影响,对点云特征的提取存在较大的不足等问题,基于现有的激光雷达点云三维物体检测网络进行了改进。通过对原始检测网络中的特征提取模块及损失函数进行优化以获得更好的检测效果,提出了改进的三维车辆目标检测网络。针对原始检测网络中特征提取时忽略了不同雷达点信息贡献程度不同的问题,改进的检测网络中通过加入特征权重学习策略,以获得更加有区分性的有效特征。此外,针对原始检测网络中存在的车辆角度约束不合理等问题,通过加入正弦函数变换,解决了车辆角度损失不合理对模型检测效果的影响。通过在公开数据集KITTI上的实验对比,结果表明了所提出的基于激光雷达的改进三维车辆目标检测网络的准确性与有效性,相对于原始检测网络,获得了更高的平均检测精度。3)针对车辆自由换道行为由于受影响因素较多而难以准确预测的问题,本文采用梯度提升决策树进行特征变换后,结合逻辑回归构建融合模型,对车辆自由换道决策行为进行建模。在真实的车辆轨迹数据集NGSIM上,通过数据滤波及筛选,构建车辆自由换道行为数据集。并且通过深入分析车辆换道决策的行为过程,挖掘更加有效的换道决策变量,以进一步提升换道决策模型的预测准确率。通过对比实验,结果表明了本文提出的融合决策模型相对于其他单一的模型能够获得更高的预测准确率。此外,新提取的碰撞时间特征在多种评价指标下均被认为是最具影响力的特征,证明了换道决策变量构建的有效性。更多还原

【Abstract】 With the development of society and economy,the concept of unmanned driving has gradually entered into daily life.Compared with manned driving systems,unmanned driving can avoid traffic accidents caused by drivers’ inattention and can liberate drivers.People are full of longing for the convenience and safety brought by unmanned driving while it has attracted the attention of researchers,but there are still many challenges in the actual landing of unmanned driving.The main problems to be solved include the accuracy of obstacle target detection in environmental perception,while the accuracy and safety of autonomous vehicle decision-making are also waiting to be solved.In this paper,the problems of low accuracy and slow speed of vehicle obstacle detection in the unmanned environment perception technology are specifically improved,while we build a model to analyze the decision-making behavior of vehicle’s free lane change.Aiming at the problems of low clustering accuracy and multiple hyperparameters in the traditional rasterized vehicle target detection method based on lidar,a density peaks clustering algorithm based on mutual neighbors with adaptive merge strategy(MNN-ADPC)is proposed based on the existing fast density peaks clustering algorithm(DPC);Aiming at the problems of insufficient point cloud feature extraction and unreasonable loss constraints in the lidar-based neural network vehicle target detection method,an improved 3D target detection network that optimizes the feature extraction module and loss terms is proposed;Aiming at the problem that vehicle’s free lane changing behavior is difficult to predict accurately due to many affected factors,the gradient boosting decision tree is applied to perform feature transformation.In combination with logistic regression,a fusion decision model is constructed to analyze the vehicle’s free lane changing behavior.The main research work and results of this article are as follows1)In the traditional rasterized vehicle target detection method based on lidar,the traditional clustering method used in obstacle grid clustering has problems such as low clustering accuracy,slow speed and too many hyperparameters.These problems severely affected the accuracy and adaptability of the clustering algorithm.Through the improvement of the initial clustering center selection and the division of points to be allocated in the existing fast density peaks clustering algorithms(DPC),density peaks clustering algorithm based on mutual neighbors with adaptive merge strategy(MNN-ADPC)was proposed.The improved clustering algorithm adopts a looser initial clustering center selection method,and combines subsequent adaptive merge strategy to solve the problem of cluster center missing selection in fixed threshold screening method.Aiming at the misclassification of the nearest neighbor partitioning strategy in original DPC,the improved clustering algorithm uses a more robust mutual neighbor partitioning strategy,which solves the problem of misclassification in embedded data.The comparative experiments results on multiple data sets show that the proposed MNN-ADPC has better accuracy and adaptability.At the same time,compared with the original DPC,the improved clustering algorithm has fewer hyperparameters,and does not require human involvement.In a real unmanned environment,the MNN-ADPC can obtain more accurate clustering results.2)Aiming at the existing lidar-based neural network vehicle target detection methods,the 3D vehicle target detection network is affected by the sparse characteristics of the point cloud,and there are major problems in the extraction of point cloud features.We improved the existing 3D object detection network based on lidar by optimizing the feature extraction module and loss function in the original detection network to obtain better detection results and an improved 3D vehicle target detection network is proposed.Aiming at the problem that different lidar points have different levels of contribution in feature extraction in the original detection network,feature weight learning strategies are added to the improved detection network to obtain more distinguished effective features.In addition,for the problems of unreasonable vehicle angle constraints in the original detection network,by adding a sine function transformation,the impact of the unreasonable vehicle angle loss on the model detection effect is solved.The comparative experiment result on the public data set KITTI shows the accuracy and effectiveness of the proposed improved 3D vehicle target detection network based on lidar.Compared with the original detection network,a higher average detection accuracy is obtained.3)Aiming at the problem that vehicle’s free lane changing behavior is difficult to predict accurately due to many affected factors,the gradient boosting decision tree is applied to perform feature transformation.In combination with logistic regression,a fusion decision model is constructed to analyze the vehicle’s free lane changing behavior.On the real vehicle trajectory dataset NGSIM,the vehicle free lane change behavior dataset is built with data filtering and screening.And through in-depth analysis of the vehicle’s lane changing decision behavior process,more effective lane change decision variables are mined to further improve the prediction accuracy of the lane change decision model.Through comparative experiments,the results show that the fusion decision model proposed in this paper can obtain higher prediction accuracy than other single models.In addition,the newly extracted collision time feature is considered to be the most influential under various evaluation indicators.The characteristics prove the effectiveness of our lane change decision variables construction.更多还原