基于GA启发式抽样的交通事件自动检测

为了提高面向不平衡数据集的交通事件检测综合性能,提出了两种基于GA启发式抽样方法的交通事件检测算法.基于GA的实例选择抽样方法(GA-IS),解决非启发式抽样方法人为设定抽样率导致的检测效果不稳定问题.基于GA的支持向量选择抽样方法(GA-SS),改善学习集数据量较大时的检测效率.实验采用新加坡AYE仿真数据库,以支持向量机作为分类器进行事件检测.结果表明,基于遗传算法实例选择抽样的检测模型检测率达到94%,平均检测时间为1.413 3 min,性能指标PI为0.157;基于遗传算法支持向量选择抽样的检测模型决策时间为4.55 s,综合性能最优,其PI为0.151;基于少数类过抽样算法(SMOTE)的检测模型决策时间为35.21 s,PI为0.329,与非启发式抽样方法相比,所提方法能有效改善面向不平衡数据集的事件检测综合性能. 

An Application of Heuristic Selection Sampling Method Based on Genetic Algorithm in Detection of Traffic Incidents

In order to improve the comprehensive performance in detection of traffic incidents for an imbalanced dataset.Two automatic incident detection (AID) algorithms based on GA-based heuristic sampling method are proposed.The method of GA-based Instance Selection (GA-IS) is proposed to settle the issue of instability caused by manual setting of sampling rate in non-heuristic sampling method.The method of GA-based Support vectors Selection (GA-SS) is proposed to improve efficiency of detection under a condition of large learning datasets.In a case study, a simulation database of Ayer Rajah Expressway (AYE) in Singapore is used, and support vector machine (SVM) is adopted as a classifier to detect incidents.The results show that the detection rate in GA-IS SVM AID model is 94%, the average time to detect incidents is 1.413 3 min, and the performance index (PI) is 0.157.Meanwhile, the decision time in GA-SS SVM AID model is 4.55 s, and the PI is 0.151.The decision time in SMOTE SVM AID model is 35.21 s, and the PI is 0.329.Compared with SMOTE, the proposed methods can provide better comprehensive performance in detection of traffic incident for imbalanced datasets.