基于轨迹数据的危险驾驶行为识别方法

Risky Driving Behavior Recognition Based on Trajectory Data

连续的跟驰行为和换道行为是驾驶行为的主要构成部分,对交通拥挤和交通事故有着重要影响。通过无人机视频拍摄和图像处理方式,提取了曹安公路沿线的2个交叉路口间正常交通流状态下共600条多车高精度轨迹数据。首先,考虑车辆类型对驾驶行为产生直接的影响,分析了大车和小车的车辆轨迹特征变量分布的差异性,包括速度、加速度、碰撞时间倒数、车头时距等,在标记危险驾驶行为的过程中考虑车辆类型的影响。其次,针对不同的车辆类型,利用修正碰撞裕度对跟驰行为和换道行为进行风险性评估,将其划分为安全型和风险型。根据风险型行为发生的顺序以及持续时间,评估驾驶人的整体驾驶状态是否危险,作为危险驾驶行为识别的样本标记。分别利用离散小波变换和统计方法提取车辆轨迹的关键特征参数,为了提高模型识别效率,将关键特征参数进行排序,从而确定最优判别指标;最后,利用轻量梯度提升机(LGBM)算法对危险驾驶行为进行识别,并与随机森林、多层感知器、支持向量机等算法在精度上进行比较。研究结果表明:在上述研究条件下,LGBM算法对危险驾驶行为的理论识别率最高可达93.62%,可以实现基于机器学习算法的危险驾驶行为的高精度自动识别,该结果对于智能驾驶辅助系统的设计、道路交通安全决策的制定具有显著的意义。

Risky driving behavior is categorized into successive car-following and lane-change maneuvers,which have an influence on traffic congestion and accidents.Approximately 600 multi-vehicle high-precision trajectory data were extracted from a video filmed using an unmanned aerial vehicle(UAV)between two intersections along Cao’an highway in normal traffic flow using image processing.First,the impact of vehicle type on the driving behavior was considered.The difference in the distribution of trajectory characteristics,including speed,acceleration,inversed time to collision(ITTC),and time to headway(THW)between cars and trucks,was analyzed.The impact of vehicle type was considered when labeling risky driving behavior.Then,modified margin to collision(MMTC)was used to evaluate the risk of car-following and lanechange maneuvers for different types of vehicles,which could be classified into safe and risky.According to the order of occurrence and duration of risky maneuvers,the overall driving behavior of drivers was evaluated as safe and risky,which were the sample labels for recognizing risky driving behaviors.The key feature parameters of vehicle trajectory were extracted using discrete wavelet transform(DWT)and statistical methods.To improve the model recognition efficiency,the key feature parameters were sorted to determine the optimal recognition indicators.Furthermore,this paper used the light gradient boosting machine(LGBM)algorithm to identify risky driving behavior.Random forest,multi-layer perception,and support vector machine algorithms were adopted to be compared with LGBM.It was found that the theoretical highest recognition accuracy of risky driving behavior is 93.62%by LGBM under the conditions of this article.This study can realize the high-precision automatic identification of risky driving behavior based on machine-learning algorithms,which is important for the design of intelligent driving assistance systems and decision-making for road traffic safety.