多源数据融合的城市区域时变停车需求识别方法

Recognition and Monitoring of Parking in Urban Region Based on Multisource Data

城市时变停车需求的准确感知是制定停车管理政策的前提条件之一,然而由于难以获取城市所有停车场进出车辆数据,且部分老旧停车场尚未实现信息化管理,以及违章停车的存在,导致城市区域时变停车需求识别困难。本研究利用城市已有交通信息采集设备,以卡口车牌识别数据、停车场进出车辆数据和出租车GPS数据为基础,通过训练模型区分停车与行车在前后两次卡口数据中的特征差异来识别车辆停车行为,从而实现城市区域时变停车需求识别。首先,结合493个停车场的进出车辆数据和3649辆出租车的GPS数据确定车辆行停状态,得到846204个停车样本和81654个行车样本,并基于合成少数类过采样技术(Synthetic Minority Oversampling Technique,SMOTE)+最近邻(Edited Nearest Neighbor,ENN)算法对不均衡行停数据集进行组合采样;其次,利用梯度提升决策树(Gradient Boosting Decision Tree,GBDT)来识别车辆的停车行为,结果表明本文建立的停车行为识别模型准确率为93.1%,精确率为99.6%,召回率为86.5%,识别效果优于其他传统方法;最后,应用该方法对某城市区域时变停车量进行监测,并利用该市41个停车场进出车辆数据进行验证,其中可通过本文算法识别的停车量为96%。研究成果可为城市区域时变停车需求感知与管理提供数据支持。

The accurate perception of time-varying parking demand in an urban region directly affects the effect of traffic management measures.However,it is difficult to obtain the parking records of all parking lots,and information management is not implemented in some old parking lots.Moreover,the illegal parking also results in the challenging of accurate parking demand recognition in urban areas.In this study,data obtained from existing traffic information acquisition equipment,including license plate recognition,parking lot,and taxi global positioning system(GPS),are used to distinguish the feature differences between parking and driving in two adjacent license plate recognition records using a training model to recognize vehicle parking status such that time-varying parking volume monitoring can be realized in urban regions.First,based on parking records of 493 parking lots and GPS data of 3649 taxis,vehicle parking status is verified;subsequently,846204 parking records and 81654 driving records are obtained,and the imbalance problem is solved using the synthetic minority oversampling technique and the edited nearest neighbor data cleaning strategy.Second,a parking status recognition model is established using gradient boosting decision tree.The results show that the accuracy,precision,and recall rates of the parking status recognition are 93.1%,99.6%,and86.5%,respectively,indicating its better performance compared with traditional methods.Finally,the method is applied to monitor the time-varying parking volume in a city and then verified based on the parking records of 41 parking lots.The parking volume that can be recognized using this algorithm is 96%,which facilitates urban parking demand monitoring and management.