【作者】 赵宇；

【导师】 杨杰；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2017， 硕士

【摘要】 在计算机视觉领域,视频中的异常行为检测一直是一个重要的研究方向,其可以被广泛地应用于场景分析,视频摘要,监控预警等诸多场景。本文针对复杂环境下目标重叠视频里的异常行为检测和目标分离视频里的异常行为检测这两个不同的问题进行了研究和探索,主要创新点包括:底层特征的设计和提取,高层视频表达的构建,建模对象的选择。围绕这几个创新点,本文提出了如下三种新的异常行为检测算法。1.基于拉普拉斯稀疏编码的异常行为检测算法。针对发生目标重叠的监控视频,由于此时对目标进行检测和跟踪比较困难,本文采用了基于局部视频特征结合拉普拉斯稀疏编码方法的检测算法进行异常行为的检测。首先,对视频在三维空间上进行局部特征抽取,然后在训练集上使用随机选择的局部特征进行字典的学习,接着将每一个视频的所有局部特征在字典上进行拉普拉斯稀疏编码,即将局部特征进行再表达,得到视频的高层特征表达,然后将每一个视频的拉普拉斯稀疏编码进行池化融合,得到一个待检测视频最终的视频级特征,最后使用支持向量机对得到的视频特征进行分类,判断待检测视频是否有异常发生。在人群密集和目标重叠的视频数据集上得到的实验结果表明这个方法对全局异常行为检测和局部异常行为检测都有良好的性能。2.基于非负局部性约束线性编码的异常行为检测算法。在目标重叠的视频里,由于要先采集底层特征,然后再对特征进行编码表达,所以底层特征对视频里不同行为的辨识能力以及特征编码方法的性能这两点对异常行为检测的结果有比较大的影响。例如在编码阶段,由于编码是将底层特征在一个预先学习的过完备字典上进行重新表达,此过程会伴随着特征信息的损失,这将影响最终的异常行为检测效果。为了进一步提高异常行为检测的准确性和算法的性能,本文提出了第二种检测框架:采用融合位置信息的底层特征和非负局部性约束线性编码相结合的检测算法。位置信息的加入使得底层特征对于视频里不同行为模式之间具有更高的辨识能力,非负局部性约束线性编码使得底层特征的编码结果不但满足稀疏性,而且满足局部性,这进一步减少了编码时的信息损失,使得最终得到的视频特征更稳健,更有区分能力。在两个主流的视频数据集上的实验结果表明这个方法在目标重叠的异常行为检测问题上有着更优异的性能。3.基于运动目标轨迹的异常行为检测算法。本文使用此算法来解决目标分离条件下的异常行为检测问题。由于在道路等监控场景下,一般异常行为的发生频率很低,如果直接对异常行为进行建模和学习,那需要的监控视频时间长度是巨大的,因此本文采取对正常行为进行建模的策略。首先,对进入视场里的目标进行跟踪,得到完整的目标轨迹,然后对每一条轨迹提取特征,此特征是速度特征和空间信息特征的组合,接着在训练数据集里对得到的正常轨迹特征进行聚类,得到正常行为的模式中心,将这些中心作为这个监控场景下正常行为的代表。对于待检测的每一条轨迹,将它和所有的模式中心进行相似性比较,如果待检测轨迹的特征和模式中心至少有一个相似,则待检测目标的行为是正常的,如果都不相似,则待检测目标的行为是异常的,在多个道路监控视频里的实验结果表明此方法对于目标分离条件下的异常行为检测有良好的性能。更多还原

【Abstract】 In the domain of computer vision,the abnormal activities detection in videos has been an important research task.It can be widely applied to a lot of scenarios such as the video summary,scene analysis,and monitoring.In this paper,we respectively apply the exploration and research on the abnormal activities detection under the conditions of overlap target and separation target.And the main innovation of this paper includes the design and extraction of feature,the construction of high-level video representation,and the selection of modeling pattern.Around these innovations,this paper proposes the following three kinds of algorithms for abnormal activity detection.1.Abnormal activity detection based on Laplace sparse coding.In the target-overlap videos,it’s difficult to detect and track targets.Therefore,this paper utilizes a method based on the local video feature and Laplacian sparse coding to detect the abnormal activities.Firstly,the local video features are extracted in the three dimensional space.Secondly,the Laplacian sparse coding scheme is used to encode local video features to produce the high level video representations.Then,the pooling method such as max pooling is adopted to integrate all code coefficients of a video to one feature vector.Finally,the support vector machine is utilized to classify these feature vectors as abnormal or normal.Experiments on two datasets demonstrate the satisfactory performance of the proposed approach.2.Abnormal activity detection based on nonnegative locality-constraint linear coding.In the videos which contains many overlap targets,the discrimination of local video feature and the performance of the coding scheme have a great influence on the detection of abnormal activities.For example,because the coding scheme is a reconstruction of the local video feature based on the pre-trained over-complete codebook,the inevitable information loss of the coding phase may influence the performance of the proposed approach.In order to further improve the accuracy of abnormal activity detection,this paper proposes another detection framework via the position-based local feature and the nonnegative locality-constrained linear coding.The introduction of the position information makes the local video feature more discriminative.The nonnegative locality-constrained linear coding makes the coding coefficient meet not only sparsity but also locality,which can reduce the reconstruction error and produce more robust and descriptive video representation.Experimental results have indicated more promising performance of the new method on the abnormal activity detection in target-overlap videos.3.Abnormal activity detection based on the trajectory of moving target.For the abnormal activity detection in the target-separation videos,this paper utilizes a method based on the trajectory of moving target.Because the abnormal activity usually occurs with a low frequency in the scenarios such as the crossroad or square,the video length will be very large if we directly model and learn the abnormal pattern.Therefore,this paper utilizes the strategy which models the normal pattern.First of all,the tracking technique is applied to gain the entire trajectories from targets which enter the view field.Then,the speed feature and space information are extracted from these trajectories.After that,the clustering scheme is utilized to produce centers of the normal pattern.For each trajectory to be detected,we calculate the similarity between it and each center.If the trajectory is similar to at least one of all the centers,it is normal.If it is not similar to any center,it is abnormal.Experimental results on two surveillance videos have demonstrated that the proposed method can perform effectively in the target-separation abnormal activity detection.更多还原