基于改进ACGAN算法的车道排队车辆估计及其分类  

Lane queue vehicles estimation and classification based on improved ACGAN algorithm

作　　者：郭海锋[1] 杨宪赞 金峻臣 Guo Haifeng;Yang Xianzan;Jin Junchen(College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023;Institute of Intelligent Transportation, Enjoyor Co., Ltd, Hangzhou 310030)

机构地区：[1]浙江工业大学信息工程学院,杭州310023 [2]银江股份智慧交通研究院,杭州310030

出　　处：《高技术通讯》2020年第11期1169-1177,共9页Chinese High Technology Letters

基　　金：浙江省自然科学基金(LY20E080023);浙江省教育科学规划(2016SCG241)资助项目。

摘　　要：针对传统模型驱动的排队车辆研究中构建概率分布困难、建模繁琐等问题,结合双向长短时记忆(Bi-LSTM)网络和辅助分类器生成对抗网络(ACGAN)的特点,提出一种数据驱动的车道级排队车辆估计算法。该算法无需对交叉口空间关系建模,其生成器采用Bi-LSTM结构,以速度序列为输入,根据速度与排队车辆的时间相关性,生成最小、最大排队车辆。判别器来自ACGAN,在区分真假样本的同时实现排队车辆到拥堵等级标签的分类。同时,为避免网络训练不稳定、梯度消失的问题,舍弃原ACGAN的真假二分类任务,引入Wasserstein散度来衡量真实序列与生成序列的分布距离,并对相应的目标函数进行优化。结果表明,与其他算法相比,该算法在分类准确率方面提高了3.96%~9.62%,同时总体估计误差最小,验证了利用速度估计车道排队车辆的可行性。Considering the difficulty of constructing probability distribution and modeling in traditional model-driven queue vehicle researches,a data-driven estimation algorithm for lane-level queue vehicles combining bi-directional long short-term memory(Bi-LSTM)network and auxiliary classifier generative adversarial network(ACGAN)is proposed,in which the intersection spatial relation model is not required.The generator turns velocity sequence into queue vehicles by Bi-LSTM according to the time correlation between them.The discriminator comes from ACGAN,categorizing the queue vehicles to congestion level labels while distinguishing true and false samples.To avoid training instability and gradient disappearance,the original ACGAN’s true-false binary classification task is abandoned,and Wasserstein divergence is introduced to measure the distribution distance between the true and false sequences,with the corresponding objective function optimized.The results show that compared with other algorithms,the classification accuracy of the proposed algorithm increases by 3.96%-9.62%with the minimal overall error,which verifies the feasibility of estimating lane queue vehicles by velocity.

关 键 词：辅助分类器生成对抗网络(ACGAN) 双向长短时记忆(Bi-LSTM) Wasserstein散度 车道级排队车辆估计 分类 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]