严妍
恶劣天气对于高速公路交通安全影响显著，对于自动驾驶车辆而言，由于恶劣天气所造成的行车环境会显著影响车辆的制动性能，从而会对交通流的安全性产生影响。传统交通流的风险评估主要依赖宏观交通流数据，在自动驾驶交通流环境下，由于车辆轨迹数据的易获得性，从微观层面进行交通风险评估具备了可行性，数据的多源性使得管理者提前预测天气对交通流的风险影响并及时做出响应成为可能。因此本文以恶劣天气下高速公路自动驾驶交通流为研究对象，建立了跟驰过程的风险评估指标。并基于ACC车辆的跟驰模型，构建仿真环境，获取不同天气条件下的自动驾驶交通流数据，建立贝叶斯网络模型评估车流风险状态，构建恶劣天气下高速公路自动驾驶交通风险预警系统。本文研究为恶劣天气下高速公路自动驾驶交通流风险预警提供有效的研究思路和方法，为未来自动驾驶车辆环境下的高速公路交通风险管控奠定了研究基础，有利于建立恶劣天气条件下的高速公路自动驾驶交通流交通管理和控制策略，提高高速公路自动驾驶交通流在恶劣天气的通行能力和运行的安全性。
首先，运用上海市自然驾驶数据，分析恶劣天气对跟驰行为的影响。研究发现天气对微观跟驰行为有显著的影响，驾驶人倾向于在恶劣天气下保持更大的跟驰间距，和更低的跟驰车速。这都说明了天气的影响会使驾驶人采取更为谨慎的跟驰行为。基于上海自然驾驶数据，还分别对晴天、雨天以及大雾天气分别标定并验证了四种典型的跟驰模型。通过对比不同天气下的跟驰参数特征以及跟驰模型标定结果，发现天气因素对跟驰参数的影响主要体现在期望跟驰间距和期望跟驰车速两方面。自动驾驶车辆通过车辆的传感器、雷达检测等装置，能够实时地根据车辆的行驶环境调节行驶状态。在未来自动驾驶环境下，恶劣天气对交通流造成的安全影响完全可以通过主动安全管理手段降至最低。因此对恶劣天气下的高速公路自动驾驶交通流风险预警技术展开研究非常必要，提前识别可能的交通风险并做出预警能够帮助管理者及时制定管控策略，降低交通流风险。
之后，提出了基于风险场理论的跟驰过程风险评估指标：风险斥力，它是一个由微观车辆轨迹计算得到的指标。通过运用美国NGSIM数据和100-Car数据库中的交通事故数据与传统风险评估指标TTC对比，论证了风险斥力指标不仅能够定量地描述驾驶人在跟驰过程中所感知的风险，同时其性能优于TTC指标，该指标还可以通过调整安全间距的中的部分参数，区分自动驾驶车辆和人工驾驶车辆，从而可以评估自动驾驶车辆在跟驰过程中面临的风险，使得自动驾驶车辆某种程度上具有人的“感知能力”，该指标能够在未来自动驾驶交通流的环境中，作为管理者制定安全管理策略的理论基础。
运用风险斥力指标作为风险评判的标准，提出了基于贝叶斯网络学习的恶劣天气下高速公路自动驾驶交通流风险评估模型，贝叶斯网络学习算法能够对自变量间存在因果关系数据进行很好的解释。使用SUMO仿真软件，考虑天气，流量，车速，车头间距，车头时距，速度差等风险相关的变量，建立仿真模型并获取了不同天气条件下的自动驾驶车辆的交通流轨迹数据。使用轨迹数据建立交通风险贝叶斯网络模型，并对模型预测精度进行检验。模型检验的结果显示，使用15s集计后的轨迹数据，模型预测的精度最好。预测准确率达到78.08%，ROC曲线AUC值达到0.808。说明模型精度较好。
最后，基于风险斥力指标和贝叶斯交通风险网络模型构建了恶劣天气下高速公路自动驾驶交通风险预警系统。预警系统包括行车风险预警和交通流风险预警两个部分。行车风险预警中使用事故数据中风险斥力的分布特征，将行车风险划分为3个等级。该预警功能旨在对行驶中的车辆的风险状态进行实时检测，并实时对高风险状态的车辆发出预警，并给出自动驾驶车辆的安全车距建议值，可以通过实时调整自动驾驶车辆速度配置文件，从而影响车辆行为保持安全车距。对于交通流风险预警系统，使用仿真数据，以路段中所有车辆的风险值均值作为交通流风险评估指标，对仿真数据中各个区段的交通流风险值进行聚类，将交通流风险划分为3个等级。其功能在于当下游路段出现恶劣天气时，将实时对上游车流可能的风险状态进行预测，并对高风险情况发出预警。
关键词：恶劣天气，自动驾驶，交通风险评估，风险预警
ABSTRACT
Adverse weather has a significant impact on highway traffic safety. For autonomous vehicles, the driving environment caused by adverse weather will significantly affect the braking performance of vehicles, which will have an impact on the safety of traffic flow. The risk assessment method of traditional traffic flow mainly depends on the macro traffic flow data. In the environment of autonomous vehicle traffic flow, due to the easy access of vehicle trajectory data, it is feasible to carry out traffic risk assessment from the micro level. The multi-source data makes it possible for managers to predict the risk on traffic flow under adverse weather in advance and make timely response. Therefore, this paper takes the highway autonomous vehicle traffic flow under adverse weather as the research object, and establishes the risk assessment indicator of the car-following process. Based on the car-following model of ACC vehicles, the simulation model is constructed to obtain the data of autonomous vehicle traffic flow under different weather conditions, the Bayesian network model is established to evaluate the risk status of traffic flow, and the risk early-warning system of autonomous vehicle traffic flow under adverse weather is constructed. This study provides effective methods for the risk early-warning of freeway autonomous vehicle traffic flow in adverse weather and lays a research foundation for freeway traffic risk management and control under the environment of autonomous vehicles in future, which is conducive to the establishment of freeway autonomous vehicle traffic flow management and control strategies under adverse weather conditions, and improves the traffic capacity and operation safety under adverse weather.
First of all, using the natural driving data of Shanghai, this paper analyzes the influence of adverse weather on the car-following behavior. It is found that weather has a significant impact on micro car-following behavior, and drivers tend to maintain a larger space headway and a lower speed under adverse weather. It shows that the impact of the weather will make the driver be more cautious. Based on the natural driving data of Shanghai, four typical car-following models are calibrated and verified on sunny, rainy and foggy days respectively. By comparing the characteristics of the car-following parameters and the calibration results of the car-following model under different weather conditions, it is found that the influence of weather factors on the car-following parameters is mainly reflected in the expected space headway and the expected speed. The autonomous vehicle can adjust the driving behavior in real time according to the driving environment of the vehicle through data from sensors, radar detection devices and other devices. In the future autonomous driving environment, the safety impact of adverse weather on traffic flow can be minimized through active safety management. Therefore, it is very necessary to study the risk early-warning technology of autonomous vehicle traffic flow under severe weather. Identifying the possible traffic risk in advance and making early warning can help managers to formulate management and control strategies and reduce the traffic flow risk in time.
Then, the risk evaluation indicator of the car-following process based on the risk field theory is put forward, risk repulsion, which is an indicator calculated from the micro vehicle trajectory. Using the traffic accident data of 100-Car database and NGSIM data, by comparing risk repulsion with the traditional risk assessment indicator, TTC, this paper proves that the risk repulsion can not only describe the risk perceived by drivers in the car following process, but also has better performance than TTC. The risk repulsion indicator can also distinguish autonomous vehicles and manual driving by adjusting some parameters in the safety distance. The risk repulsion indicator can assess the risk faced by an autonomous vehicle in the car-following process, so that the autonomous vehicle to some extent has the "perception ability" of people, which can be used as the theoretical basis for managers to formulate safety management strategies in the future environment of autonomous traffic flow.
Based on the risk repulsion indicator, this paper puts forward the risk evaluation model of freeway automatic driving traffic flow in adverse weather using the Bayesian network learning algorithm. Bayesian network learning algorithm can well explain the data which exist causal relationship between variables. Using SUMO simulation software, the simulation model is established and the traffic flow trajectory data of autonomous vehicles under different weather conditions are obtained. The Bayesian network model of traffic risk is established by using trajectory data, and the prediction accuracy of the model is tested. The results of model test show that the prediction accuracy of the model is the best when the trajectory data of 15s is used. The prediction accuracy is 78.08%, and AUC of ROC curve is 0.808. It shows that the accuracy of the model is good.
Finally, based on the risk repulsion indicator and Bayesian traffic risk network model, the risk early-warning system of freeway autonomous vehicle traffic flow in adverse weather is constructed. Early-warning system includes two parts: risk early warning for car following process and risk early warning for traffic flow. In the early warning for car following process, using the distribution characteristics of risk repulsion in accident data, the driving risk is divided into three levels. The purpose of this early warning function is to detect the risk status of the vehicle in real time, and give an early warning to the vehicle when it is in high risk status, and recommend the safe distance of the autonomous vehicle. The speed profile of the autonomous vehicle can be adjusted in real time, so as to affect the vehicle behavior and maintain the safe distance. For the traffic flow risk early warning system, using the simulation data, taking the average risk value of all vehicles in the road section as the traffic flow risk assessment indicator, clustering the traffic flow risk value of each section in the simulation data, and dividing the traffic flow risk into three levels. Its function is to predict the possible risk state of the upstream traffic flow in real time and give an early warning of the high-risk situation when there is adverse weather in the downstream section.
Key Words: Adverse weather, Autonomous vehicle, Traffic risk assessment, Risk early-warning.
