【作者】 郭剑；

【导师】 李芬兰；

【作者基本信息】 汕头大学 ， 信号与信息处理， 2010， 硕士

【摘要】 基于混合高斯隐马尔科夫模型(GHMM)的轨迹识别中,关键的问题是运动对象轨迹的有效获取,以及基于这些轨迹的健壮表示和建模,然后根据模型来识别轨迹。本文以视频运动对象轨迹识别为技术目标,从低层视频分析技术开始,系统研究了轨迹识别框架下的目标检测、跟踪和轨迹建模与识别算法,重点研究了轨迹建模与识别,采用混合高斯隐马尔科夫模型对获取的轨迹进行建模。为了降低特征空间的维数,减少计算量,把轨迹用参数函数的形式表示成一条曲线,在曲率变化最大的点处分割轨迹为一些相似的极小单元,称其为子轨迹,用它们的主成分分析(PCA)系数来表示该轨迹,用PCA系数来训练混合高斯隐马尔科夫模型,并且对运动目标轨迹进行识别。本文分析视频来自CAVIAR项目,是从固定摄像头获取的,对顾客在超市门口活动的轨迹进行识别,从而得知顾客的具体活动。我们把顾客的活动分为六类:从左到右,从右到左,从左边进去,从右边进去,从左边出来和从右边出来。具体实现过程分为以下几步:首先实现场景中人的检测和跟踪,并提取轨迹;然后用均值聚类法对轨迹聚类,把轨迹分为六类,对每类轨迹进行PCA系数表示;最后对每类轨迹建立GHMM,通过计算每个轨迹的识别率来验证模型的有效性。实验结果表明,基于PCA系数表示的GHMM有着很好的轨迹识别能力。通过与整个轨迹点为输入向量的GHMM识别比较,该方法不仅降低了特征空间维数,还提高了识别率。更多还原

【Abstract】 In the field of trajectory recognition based Gaussian Mixture Hidden Markov Model (GHMM), the key issues is obtaining trajectory of moving objects, robust presenting and modeling of the trajectory. Furthermore, we analyze activities by the models. Given technical objectives of trajectory recognition, this paper investigates systemically many Algorithms, following the framework of trajectory recognition, such as moving objective detecting, Tracking and trajectory modeling etc. We concentrated on trajectory modeling. We model trajectory by Gaussian Mixture Hidden Markov Model. Before modeling, trajectories are segmented at points of change in curvature, and the subtrajectories are represented by their principal component analysis (PCA) coefficients, to reduce the feature space dimension. Then we present a framework to robustly estimate the multivariate probability density function based on PCA coefficients of the subtrajectories using GHMMs.The Experimental video is from CAVIAR project in the paper, which is obtained from a fixed camera. We can analysis activity of customers’coming and going by trajectory recognition. The customers’activities include walking from left to right, walking from right to left, entering from the left, entering from the right, out from the left and out from the right. This process is divided into three stages. First, we obtain the trajectory through detecting and tracking. Second, we cluster the trajectory into six categories by K-means clustering, and PCA reducing dimensionality is used to process trajectory; Finally, we will establish GHMM for each trajectory, and Validate the validity of the model by calculating the recognition rate of each trajectory. Experiments demonstrate the superiority of our proposed HMM-based scheme using PCA coefficients of subtrajectories in comparison with other techniques in the literature.更多还原