基于二次谱聚类和HMM-RF混合模型的车辆行为识别方法研究  

Vehicle Behavior Recognition Method Based on Quadratic Spectral Clustering and HMM-RF Hybrid Model

作　　者：范菁[1,2] 阮体洪 吴佳敏[1] 董天阳[1,2] 

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]浙江省可视媒体智能处理技术研究重点实验室,杭州310012

出　　处：《计算机科学》2016年第5期288-293,共6页Computer Science

基　　金：浙江省重大科技专项重大工业项目(2013C01112)资助

摘　　要：从高速交通监控视频中提取的车辆轨迹数据可以用于分析和识别车辆行为。由于从高速监控视频中提取的车辆轨迹中只有少量的变道、超车等类型轨迹,采用经典的最长公共子串(LCSS)相似度和谱聚类等算法无法有效地区分轨迹数据中所有类型的轨迹;此外,在车辆行为识别方面,常用的隐马尔科夫(HMM)轨迹模型忽略了负样本的影响,且仅用最大似然值进行分类,存在较高的误识别率。为了解决这些问题,分析和研究了高速监控视频中车辆轨迹数据的特点,提出了一种基于二次谱聚类和HMM-RF混合模型的车辆行为识别方法。该方法利用轨迹曲率来识别具有曲线轨迹特征的超车轨迹,利用倾角相似度和谱聚类算法来识别非曲线轨迹中的变道轨迹,并将得到的所有聚类簇用LCSS和谱聚类算法进行再聚类,从而有效地区分超车、变道以及直行轨迹等。在进行车辆行为识别时,该方法通过将不同HMM模型的多维概率输出作为随机森林RF模型的输入来识别多类型轨迹以替代最大似然值分类,提高了行为识别的正确率。为了验证方法的有效性,在不同数据集下进行实验,结果表明轨迹聚类的平均准确率为96%,而行为识别的平均准确率是89.3%,算法具有较高的准确率和鲁棒性。The vehicle trajectory extracted from highway surveillance system can be used to analyze and recognize vehicle behavior.Due to a small amount of abnormal trajectory,such as change lanes and overtaking,the classic spectral clustering with longest common sub-sequence（LCSS）can＇t effectively distinguish all kinds of trajectory.In addition,the popular HMM trajectory model ignores the negative impact of the samples and only classifies them by maximum likelihood value to cause a higher rate of false recognition in vehicle behavior recognition.In order to address these issues,according the characteristics of highway vehicle trajectory,we proposed a vehicle trajectory recognition method based on quadratic spectral clustering and HMM-RF hybrid model.Firstly,the trajectory curvature is calculated to distinguish overtaking by curved characteristics,and then lane changes trajectory is distinguished by spectral clustering with inclination similarity in non-curve trajectory.Secondly,all the sub-clusters are clustered by spectral clustering with LCSS again,which can effectively distinguish overtaking,changing lanes and normal trajectory.We made the output of HMM model,the different dimension of probabilities as an input for random forest model to improve the precision of behavior recognition.We did experiments under different data sets to verify the effectiveness of the method.The average accuracy rate of trajectory clustering can achieve 96%,and the average accuracy rate of behavior recognition can reach to 89.3%,so the algorithm has higher accuracy and robustness.

关 键 词：轨迹聚类 车辆行为识别 二次谱聚类 HMM-RF混合模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]