【作者】 王伟；

【导师】 漆巍巍；

【作者基本信息】 华南理工大学 ， 交通运输管理与规划， 2022， 硕士

【摘要】 高速公路合流区作为一种连续流性质的交通设施,一方面提高了路网的通达度和路网的运行效率,另外一方面也造成了多股车流汇集,产生了大量的换道冲突。研究表明,高速公路上有近四分之三的交通事故直接或间接与合流区内频繁的轨迹交叉和换道冲突有关。因此高速公路合流区的运行质量与高速公路路网的运行效率和运行安全性存在着密切的联系。利用交通冲突技术进行高速公路合流区的安全分析依赖于大量的车辆轨迹数据。人工的车辆轨迹数据采集方法费时费力,且无法提供高精度的数据;GPS、雷达等设备虽然能提供精度较高的轨迹数据,但是需要对特定道路进行改造,成本较高,可实施性较差。鉴于计算机视觉技术的高速发展,利用目标检测算法和目标追踪算法实现高速公路合流区车辆轨迹数据自动化采集和重构,能有效地提高车辆轨迹数据采集的效率和精度。以此为背景,本文基于深度学习算法实现了车辆轨迹的识别和重构。在此基础上,分析了高速公路合流区的交通流特性和交通冲突特性,构建了高速公路合流区换道冲突序列模型。最后,基于Light GBM实现了高速公路合流区换道冲突的自动判别。本文的重点研究内容总结如下:(1)为了高精度且自动化地采集高速公路合流区车辆轨迹数据,基于目标识别算法YOLOV5和目标追踪算法Deep SORT,实现了车辆位置的检测和车辆轨迹的追踪。在此基础上,利用速度和加速度公式,实现了高速公路合流区的车辆轨迹参数自动采集。最后,基于海量的标定数据,以平均精度、查准率和查全率为指标,验证了车辆位置检测和轨迹追踪的有效性。(2)针对高速公路合流区车辆轨迹参数中存在异常值和随机噪声的特点,分别选取EDM算法,EEDM算法和CEEDMAN算法进行高速公路合流区车辆轨迹的重构运算,输出了车辆轨迹重构前后对比图和对比表。对比三种算法的车辆轨迹重构结果,甄选出CEEDMAN算法作为高速公路合流区车辆轨迹重构的推荐方法。(3)根据高速公路合流区的几何特性和车流特性,论证了冲突时间差(Time Difference to Collision,TDTC)作为高速公路合流区换道冲突指标的有效性。为了表征高速公路合流区换道冲突的动态特征,以TDTC为指标,构建了高速公路合流区换道冲突序列模型。同时,利用重构后的高速公路合流区车辆轨迹,以数据置信度和数据平衡为指标,验证了高速公路合流区换道冲突序列模型的有效性。(4)以高速公路合流区换道冲突序列模型所计算出的TDTC值为基础,引入Light GBM算法,实现了高速公路合流区换道冲突的自动判别。经过训练,高速公路合流区的换道冲突自动判别模型的总体正确率达到了91%以上,其在查全率和查准率这两个方面均优于随机森林、Ada Boost、XGBoost、决策树以及K近邻等几种模型。最后,以特征重要度为指标,证明了基于Light GBM的合流区换道冲突自动判别模型具备较强的可解释性和有效性。综上所述,本文通过引入深度学习算法和标定海量的轨迹样本,实现了高速公路合流区车辆轨迹参数的自动提取和重构,为车辆轨迹数据的采集提供了一个较好的解决方案。同时,本文提出的高速公路合流区车辆换道冲突序列模型和换道冲突自动判别算法,能够为高速公路合流区的安全评价与优化设计提供理论支撑。更多还原

【Abstract】 As a traffic facility of continuous flow nature,highway confluence area improves the accessibility and operational efficiency of the road network on the one hand,and on the other hand,they cause the convergence of multiple streams of traffic and generate many lane change conflicts.Research shows that nearly three-quarters of traffic accidents on highways are directly or indirectly related to frequent track crossings and lane change conflicts in confluence area.Therefore,the operation quality of the highway confluence area is closely related to the operation efficiency and safety of the highway network.The safety analysis of highway confluence area using traffic conflict technology relies on a large amount of vehicle trajectory data.Traditional trajectory data collection methods are time-consuming and laborious and cannot provide high-precision trajectory data;GPS,radar and other equipment can provide high-precision trajectory data,but need to modify specific roads,which is costly and poorly implementable.In view of the rapid development of computer vision technology,the use of target detection algorithm and target tracking algorithm to achieve automated vehicle trajectory data acquisition and reconstruction can effectively improve the efficiency and accuracy of vehicle trajectory data acquisition.In summary,this thesis realizes vehicle trajectory recognition and reconstruction based on deep learning algorithm.On this basis,the traffic flow characteristics and traffic conflict characteristics of the highway confluence area are analyzed,and a sequence model of lane change conflict in the highway confluence area is constructed.Finally,based on Light GBM,the automatic recognition of lane change conflicts in highway confluence area is realized.This thesis focuses on the study of the conflict characteristics of lane change and traffic flow characteristics within the highway confluence area,to provide a basis and support for the safety evaluation and design of the highway confluence area.The key research contents of this thesis are shown as follows.(1)In order to collect vehicle trajectory data in the highway confluence area with high accuracy and automation,the detection of vehicle position and tracking of vehicle trajectory are realized based on the target recognition algorithm YOLOV5 and the target tracking algorithm Deep SORT.On this basis,the vehicle trajectory parameters of the highway confluence area are automatically collected using the velocity and acceleration formulas.Finally,the effectiveness of vehicle position detection and trajectory tracking is verified based on a large amount of calibration data and with the average accuracy,accuracy and completeness of the check as indicators.(2)For the characteristics of outliers and random noise in the vehicle trajectory parameters in the highway confluence area,the EDM algorithm,EEDM algorithm and CEEDMAN algorithm are selected to reconstruct the vehicle trajectory in the highway confluence area,and the before-and-after vehicle trajectory reconstruction comparison chart and comparison table are output.Comparing the reconstruction results of the three algorithms,it is found that the CEEDMAN algorithm has the best effect on the reconstruction of vehicle trajectories in the highway confluence area.(3)The effectiveness of Time Difference to Collision(TDTC)as the indicator of lane change conflict in the confluence area is demonstrated based on the geometric characteristics and traffic flow characteristics of the confluence area.In order to characterize the dynamic characteristics of the conflict,the TDTC is used as an indicator to construct a sequential model of the highway confluence area lane change conflict.At the same time,the validity of the highway confluence area conflict sequence model was verified by using the reconstructed vehicle trajectories in the confluence area with data confidence and data balance as indicators.(4)Based on the TDTC values calculated by the sequence model of highway confluence area lane change conflicts,the Light GBM algorithm is introduced to achieve automatic discrimination of highway confluence area lane change conflicts.After training,the overall correct rate of the automatic discriminative model for highway confluence area lane change conflicts reaches 91%,and it outperforms several models such as random forest,Ada Boost,XGBoost,decision tree,and K-nearest neighbor in both the check-all rate and check-accuracy rate.Finally,using the feature importance as the index,the automatic discriminative model based on Light GBM for confluence area lane change conflicts is proved to be highly interpretable and effective.In summary,this thesis realizes the automatic extraction and reconstruction of vehicle trajectory parameters in the highway confluence area,and provides a new solution for the collection of vehicle trajectory data.At the same time,the highway confluence area lane change conflict sequence and automatic lane change conflict discrimination model proposed in this thesis can provide a basis for the safety evaluation and design of highway confluence area.更多还原