基于车辆行驶轨迹的道路不良驾驶行为实时辨识方法

为了提高道路交通安全主动防控能力, 以小汽车行驶轨迹数据为研究对象, 研究了不良驾驶行为的实时辨识问题; 基于无人机拍摄交通流视频提取海量车辆行驶轨迹数据; 提出了应用风险度量方法量化典型不良驾驶行为的理论; 使用大样本统计分布方法确定不良驾驶行为的特征参数阈值; 建立了结合交通环境信息的不良驾驶行为谱, 计算了不良驾驶行为谱特征值; 以车辆不良驾驶行为谱特征值为依据标定不良车辆样本; 以部分驾驶行为谱参数为输入, 使用不平衡类提升的人工智能算法建立了不良驾驶行为辨识模型; 为了验证方法的有效性, 使用无人机交通视频采集了上海市的车辆行驶轨迹数据, 分析了小汽车不良跟驰行为特征。分析结果表明: 使用四分位差法得到不良跟驰特征参数的阈值为0.19 s-1, 大部分样本处于正常跟驰状态, 约2%样本处于不良跟驰状态; 基于每辆车行驶轨迹中正常跟驰状态和不良跟驰状态的比例, 使用95%分位数将8 917 veh小汽车样本划分为不良跟驰车辆445 veh与正常跟驰车辆8 472 veh; 不平衡类提升算法CUSBoost辨识不良跟驰车辆达到了94.4%的召回率和85.9%的精确率, 平衡分数和精确率-召回率曲线下的面积为所有算法中最高。可见, 不良驾驶行为谱作为一种客观的不良驾驶行为量化表达方法, 与人工智能方法结合可以生成海量的不良驾驶行为谱库; 不平衡类提升算法可以解决不良驾驶行为数据的不平衡问题, 与常规算法相比具有更好的不良驾驶行为辨识能力。 

Real-time identification method of abnormal road driving behavior based on vehicle driving trajectory

To improve the active prevention and control of road traffic safety, the driving trajectory data of cars were used as the research object and the real-time identification problem of abnormal driving behaviors was studied. A massive amount of vehicle driving trajectory data was extracted based on an unmanned aerial vehicle recorded traffic video. The theory of applying the measurement of risk to quantify the typical abnormal driving behaviors was proposed. A large-sample statistical distribution method was used to determine the characteristic value thresholds for abnormal driving behaviors. The abnormal driving behavior spectrum was established and combined with the traffic environment information. The characteristic value of the abnormal driving behavior spectrum was calculated. Based on the characteristic value of the vehicle's abnormal driving behavior spectrum, the abnormal vehicle samples were labeled. The driving behavior spectrum parameters were partially used as input and imbalanced class boosting artificial intelligence algorithms were used to establish an abnormal driving behavior identification model. To verify the effectiveness of the method, vehicle driving trajectory data were collected using an unmanned aerial vehicle traffic video in Shanghai, and the characteristics of car abnormal car-following behavior were analyzed. Research result shows that the threshold of the abnormal car-following characteristic parameter is 0.19 s-1 using the interquartile range method, most of the samples are in the normal car-following state, and about 2% of the samples are in the abnormal car-following state. Based on the ratio of the normal car-following state and the abnormal car-following state in the driving trajectory of each car, a 95% percentile is used to divide the examples of 8 917 cars into 445 abnormal car-following cars and 8 472 normal car-following cars. The imbalanced class boosting algorithm CUSBoost achieves a recall rate of 94.4% and a precision rate of 85.9% in identifying abnormal car-following vehicles. Its balanced score and the area under the precision-recall curve are the highest among all algorithms. As an objective and quantitative expression method of abnormal driving behavior, the abnormal driving behavior spectrum can be combined with the artificial intelligence method to generate a massive library of abnormal driving behaviors. The imbalanced class boosting algorithm can address the imbalance problem of abnormal driving behavior data and has a better ability to identify abnormal driving behaviors compared to conventional algorithms.