无信号环形交叉口机非冲突机器学习预测方法

为高效精确地预测无信号环形交叉口机动车与非机动车的交通冲突,提出了基于遗传算法优化的BP神经网络(GA-BP)和支持向量回归(SVR)的组合预测模型(SVR-GA-BP)。通过无人机采集混合交通流高清视频,利用视频识别软件Tracker提取机非交通冲突轨迹数据,以距离碰撞时间(Time to Collision, TTC)为判别指标,确定机非冲突严重程度。基于偏相关性分析确定交通量、平均速度、大车比例等为机非交通冲突的显著影响因素,选取均方根误差(RMSE)、平均绝对误差(MAE)等五种评价指标对SVR模型、BP神经网络、SVR-GA-BP模型的预测值进行精度分析。结果表明,组合模型在一般冲突预测中精度为97.1%,相比SVR和BP神经网络分别提高6.9%和2.5%,在严重冲突预测中精度为96.1%,相比SVR和BP神经网络分别提高7.3%和5.1%。可见SVR-GA-BP组合模型能够有效预测无信号环形交叉口的机非冲突且精度最高,可为同类型交叉口的安全评价提供借鉴。

Traffic Conflict Prediction of Motorized and Non-motorized Vehicles at Unsignalized Roundabout Using Machine Learning

In order to efficiently and accurately predict traffic conflicts between motorized and non-motorized vehicles at unsignalized roundabouts, a combined prediction model (SVR-GA-BP) based on genetic algorithm optimized BP neural network (GA-BP) and support vector machine regression (SVR) was proposed. The high-precision mixed traffic flow trajectory data was collected using drone video at an unsignalized roundabout. The video recognition software Tracker was used to extract the trajectory data of motorized and non-motorized vehicles conflicts. The time to collision (TTC) parameter was chosen as the discriminant index to determine the severity of motorized and non-motorized vehicles conflicts. Based on partial correlation analysis, the traffic volume, average speed and percentage of heavy vehicles were determined as the significant influencing factors for conflicts. Five evaluation metrics, such as root mean square error (RMSE) and mean absolute error (MAE), were selected to analyze the accuracy of the predicted values of SVR model, BP neural network and SVR-GA-BP model. The results show that the accuracy of the combined model in minor conflict prediction is 97.1 %, which is 6.9 % and 2.5 % higher than that of SVR and BP model respectively. The accuracy of the combined model in serious conflicts prediction is 96.1 %, which is 7.3 % and 5.1 % higher than that of SVR and BP model respectively. It can be seen that the SVR-GA-BP combined model can effectively predict the motorized and non-motorized vehicles traffic conflict of unsignalized roundabout with the highest accuracy, which can provide reference for the safety evaluation of the same type of intersections.