【作者】 刘洋；

【导师】 周竹萍；

【作者基本信息】 南京理工大学 ， 交通信息工程及控制， 2023， 硕士

【摘要】 无信号控制的城市路段交通场景中,行人与车辆路权模糊、交互复杂,提前对行人的过街意图做出准确判别,对于自动驾驶车辆安全平稳运行至关重要。因此,本文面向自动驾驶技术,开展对城市无信号控制路段中的行人行为预测进行研究,旨在通过分析路段行人过街的复杂社会交互作用,进一步准确预测行人未来的过街意图,提高行人过街的安全性以及行车效率,促进自动驾驶技术向高水平发展。首先,针对现有国内外公开数据集对于行人意图的注释数据较少,借助南京交通大数据研判中心的固定监控视角,本文采取两种方案自主采集研究场景所需数据,并结合自动驾驶公开数据集nu Scenes,共采集3395段行人过街样本。此外,通过构建固定监控视角与移动车载视角两种视角下的坐标系,结合视频检测算法提取行人过街行为特征参数,为行人过街社会交互特性分析与过街意图预测提供丰富的数据基础。其次,从人-人交互、人-车交互与人-环境交互三个方面,分析行人过街社会交互特性,通过构建人-人交互力、人-车交互力与人-环境交互力对所有交互对象进行联合交互建模,解决了多行人、多车辆之间的动态交互过程难以量化的问题。其中人-人交互通过构建行人群网络阈值模型分析了行人群体过街的从众效应;人-车交互基于行人与自动驾驶车辆、传统车辆交互的差异性研究,提出了Lane-Distance-Velocity模型,划分了多车道、多车辆行人过街概率区域;人-环境交互通过构建梯度提升决策树模型分析道路结构、交通特征与天气特征对行人过街的影响程度。最后,为捕捉更多信息预测行人过街意图,本文在经典隐马尔可夫模型基础上,提出依赖型隐马尔可夫模型,将交互力与行人头部朝向作为模型输入,基于行人前2.0秒观测状态估计行人过街意图,并搭建滑动时间窗,实时更新行人观测状态,滚动预测行人过街意图,输出行人未来1.0秒的过街意图序列。通过Akaike信息准则和贝叶斯信息准则验证了依赖型隐马尔可夫模型在处理行人过街意图预测问题上优于经典隐马尔可夫模型。实验结果表明:所提出的模型在行人过街意图识别任务上,精度为91.46%,提前0.5秒预测行人过街意图,精度为88.71%,提前1.0秒预测行人过街行人意图,精度为85.13%。并且对比不同模型算法,证实本文考虑行人的复杂社会交互作用能有效提高模型预测精度,最后通过分析三类行人特殊过街场景,进一步验证了行人社会交互特征建模的合理性。更多还原

【Abstract】 The midblock without traffic lights in urban,the interaction between pedestrians,vehicles and environment is complex,accurate identification of pedestrians ’ intention to cross the street in advance is crucial for the safety and stability of Autonomous Vehicles.Therefore,this thesis studies the prediction technology of pedestrian group crossing intention in urban unsignalized control midblock.The aim is to further identify the crossing intention of pedestrians by analyzing the complex social interaction of pedestrian crossing,to improve the safety and driving efficiency of pedestrian crossing,and promote the development of autonomous driving technology to a high level.Firstly,in view of the fact that the existing open-source datasets at home and abroad have less annotation data for pedestrian intentions,with the help of the overhead monitoring cameras of Nanjing Traffic Big Data Research and Judgment Center,this thesis adopts two schemes to independently collect the data required for the research scene,and combines the dataset of nu Scenes to collect 3395 pedestrian crossing samples.In addition,by constructing a coordinate system from two perspectives of the overhead monitoring perspective and mobile vehicle perspective,combined with video detection algorithm to extract pedestrian crossing behavior characteristic parameters,it provides a rich data basis for pedestrian crossing social interaction characteristics analysis and crossing intention prediction.Secondly,the social interaction characteristics of pedestrian crossing are analyzed from three aspects: pedestrian-pedestrian interaction,pedestrian-vehicle interaction and pedestrianenvironment interaction.By constructing pedestrian-pedestrian interaction force,pedestrianvehicle interaction force and pedestrian-environment interaction force,all interaction objects are jointly modeled,which solves the problem that the dynamic interaction process of multipedestrian and multi-vehicle is difficult to quantify.Among them,pedestrian-pedestrian interaction analyzes the conformity effect of pedestrian group crossing by constructing a threshold model of pedestrian network.Based on the difference research of pedestrian-vehicle interaction with Autonomous Vehicles and Human-driven vehicles,the Lane-Distance-Velocity model is proposed to divide the probability area of multi-lane and multi-vehicle pedestrian crossing.Pedestrian-environment interaction analyzes the influence of road structure,traffic characteristics and weather characteristics on pedestrian crossing by constructing Gradient Boosting Decision Tree model.Finally,in order to use more information to predict the pedestrian crossing intention,this thesis proposes a Dependent Hidden Markov Model based on the classical Hidden Markov model,estimates the pedestrian crossing intention based on the observation state of the first2.0s of the pedestrian,and builds a sliding time window to update the pedestrian observation state in real time,rolling predict the pedestrian crossing intention,and output the pedestrian crossing intention sequence of the next 1.0s.The AIC and BIC values verify that the Dependent Hidden Markov Model is superior to the classical hidden Markov model in dealing with pedestrian crossing intention prediction.The experimental results show that the accuracy of the proposed model in pedestrian crossing intention recognition task is 91.46%,the accuracy of pedestrian crossing intention prediction is 88.71% in 0.5 seconds ahead of time,and the accuracy of pedestrian crossing intention prediction is 85.13% in 1.0 seconds ahead of time.By comparing different model algorithms,it is proved that the complex social interaction of pedestrians can effectively improve the prediction accuracy of the model.By analyzing three kinds of pedestrian special crossing scenes,the rationality of pedestrian social interaction feature modeling is further verified.更多还原