摘要
为解决交通事件检测中特征稀少与样本失衡的问题,提出一种基于变分自编码器(VAE)-随机森林(RF)-分类梯度提升树(CatBoost)的交通事件检测框架。依据4条规则构建较为全面的初始特征集。使用VAE进行数据平衡。之后,采用RF算法筛选出最佳输入特征集。最后,采用CatBoost算法作为分类器检测交通事件。使用真实世界的交通数据集设计实验,并选择6个有效评价指标对实验结果进行评价。结果表明:所提出的交通事件检测框架除误报率之外,各项评价指标均取得最优结果,表明在交通事件检测方面具备优异性能。
To address the feature scarcity and sample imbalance in traffic incident detection,this paper proposes a traffic incident detection framework based on variational auto-encoder-random forest-categorical gradient boosting tree is proposed.First,a more comprehensive initial feature set is built based on four rules.Second,a variational auto-encoder is employed to balance the datasets.Then,the Random Forest algorithm is employed to filter the best input feature set.Finally,the categorical gradient boosting tree algorithm is introduced as a classifier to detect traffic incidents.Experiments are conducted by using real-world traffic datasets and six effective evaluation metrics are selected to evaluate the experimental data.Our results show the proposed framework delivers the optimal performances in all evaluation metrics except the false alarm rate,indicating its superiority in traffic incident detection.
作者
张兵
邹少权
陆春霖
陈渤文
薛运强
ZHANG Bing;ZOU Shaoquan;LU Chunlin;CHEN Bowen;XUE Yunqiang(School of Transportation Engineering,East China Jiaotong University,Nanchang 330000,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2024年第8期197-207,共11页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金地区基金项目(52162042)
江西省省教育厅一般课题(GJJ190331)。
关键词
交通事件检测
特征扩展
数据平衡
特征选择
梯度提升树
traffic incident detection
feature expansion
data balancing
feature selection
gradient boosting tree
