【作者】 冯辉；

【导师】 刘静； 姚舜；

【作者基本信息】 西安电子科技大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 异常事件的发生对人们的生产和生活都会产生极大的影响,为了降低异常事件的影响,需要对异常事件进行实时检测。现实生活中的异常事件主要分为两类:(1)静态异常,例如工业生产中出现的产品外观缺陷;(2)动态异常,例如监控视频中行人的异常行为。这两类异常事件检测都需要对视频或图像进行理解和分析。然而,随着基础设施建设和生产力的提高,异常检测需求在很多场景中激增,导致依赖人工进行异常检测的方法变得不切实际。因此,利用计算机视觉方法来代替低效繁琐的人工检查具有重要意义。目前,已经有学者提出一些方法进行缺陷和异常事件检测,但异常事件检测面临着数据难以收集,样本比例严重不均衡等问题,这些因素极大地影响了算法的异常检测效果。在本文中,首先提出基于深度聚类网络的外观异常检测方法用于识别静态异常。然后,为了建模时序信息以提高异常检测精度,提出了基于前景背景分离的视频异常事件检测方法用于识别动态异常。最后,为了建模长时序信息和人类群体特征,提出基于时空图卷积网络和群体特征的异常行为检测方法,用于识别动态异常中的人类异常行为。具体的工作分为以下三部分内容:1.提出了基于深度聚类网络的物体外观异常检测方法在静态异常检测任务中,物体表面正常区域之间通常具有较大的差异。然而,现有的外观异常检测方法将所有正常区域视为同一类别,忽略了它们之间的差异性,因此降低了模型对正常特征的建模能力。本文提出了一种基于深度聚类网络的外观异常检测方法,该方法让不同正常区域特征自适应地向各自所属的聚类中心靠拢,从而使得正常特征表示更加紧凑。为了进一步降低误报率并优化决策边界,该方法生成了大量带有伪缺陷的样本用于训练,并通过增大正常样本特征的聚类中心与伪缺陷区域特征的距离,提高模型对正常样本和异常样本之间的预测差异。本文在一个公开数据集上对多种纹理和物体样本进行了验证,实验结果表明,所提出的算法能明显提高外观异常检测的精度。2.提出了基于前景背景分离网络的视频异常事件检测方法在实践中发现,从多个时序特征对目标的进行动态识别可以提高异常检测精度,因此,视频中的时序信息对于异常检测至关重要。在动态异常检测任务中,视频前景中目标外观和运动模式的异常表现是异常事件的关键特征,因此前景的模式识别对于异常检测至关重要。然而,视频帧的前景通常表现出一定的稀疏性,许多现有方法通过视频帧的重建任务来学习正常事件的模式。这种方法会使模型重点关注视频背景重建,而忽略对前景模式的学习。为了缓解以上问题,本文提出了一种基于前景背景分离网络的视频异常事件检测方法。该方法将前景和背景的学习分为两个子任务,保证了模型对正常事件模式的建模能力。此外,该方法利用光流的轮廓信息来辅助前景背景分离,并对正常事件的短时序信息进行建模。本文在三个公开数据集上进行了实验,并获得了先进的性能,验证了前景背景分离策略对视频异常事件检测的有效性。3.提出了基于时空图卷积网络和群体特征的视频异常行为检测方法人类异常行为识别通常依赖更长的时序信息,在动态异常检测中十分普遍,这种行为对社会治安的影响非常严重。因此,研究异常行为检测具有至关重要的意义。现有的异常行为检测方法只关注个体内部的运动状态,而忽略了个体之间交互状态的影响。本文提出了一个基于时空图卷积网络和群体特征的异常行为检测方法,用于建模人类行为的长时序信息,并融入群体交互特征。首先采用时空图卷积网络自动编码器提取个体姿态和群体姿态的融合特征,接着采用深度嵌入聚类算法对正常行为特征进行软聚类,最后采用狄利克雷过程混合模型计算人体行为的异常分数。实验结果表明,所提出的方法在两个公开数据集上的AUC指标优于已有视频异常行为检测方法。更多还原

【Abstract】 The occurrence of abnormal events will have a great impact on people’s production and life.In order to reduce the impact of abnormal events,it is necessary to detect abnormal events in real time.Abnormal events in real life are mainly divided into two categories:(1)static anomalies,such as product appearance defects in industrial production;(2)dynamic anomalies,such as abnormal behavior of pedestrians in surveillance videos.Both types of anomaly detection require understanding and analysis of video or images,and the current speed and scale of data collection makes methods that rely on manual anomaly detection impractical.Therefore,it is of great significance to use computer vision methods to replace inefficient and cumbersome manual inspections.At present,some scholars have proposed some methods to detect defects and abnormal events.However,abnormal event detection faces problems such as difficult data collection and serious imbalance of sample ratio.These factors greatly affect the abnormal detection effect of the algorithm.In this thesis,an appearance anomaly detection method based on deep clustering network is proposed to identify static anomalies.Then,in order to model time sequence information and improve the accuracy of anomaly detection,a video anomaly event detection method based on foreground background separation was proposed to identify dynamic anomalies.Finally,in order to model long time series information and human population characteristics,an abnormal behavior detection method based on spatiotemporal graph convolutional network and population characteristics is proposed to identify abnormal behaviors in dynamic anomalies.The specific work is divided into the following three parts:1.A method for object appearance anomaly detection based on deep clustering network.In static anomaly detection tasks,there are large differences between normal areas on the surface of an object.However,existing appearance anomaly detection methods regard all normal regions on the object surface as the same category,ignoring the differences among them,thus reducing the model’s ability to model normal features.This thesis proposes an appearance anomaly detection method based on deep clustering network,which allows different normal regions to adaptively move closer to their respective cluster centers,thereby making the feature representation of each cluster more compact.In order to further reduce the false alarm rate and optimize the decision boundary,this method generates a large number of samples with pseudo defects for training,and increases the distance between the cluster center of normal sample features and the pseudo defect area features.Thereby improving the model’s prediction difference between normal samples and abnormal samples.In this thesis,a variety of texture and object samples are verified on a public dataset.The experimental results show that the proposed algorithm can significantly improve the accuracy of appearance anomaly detection.2.A method for video anomaly event detection based on foreground and background separation network.In practice,it is found that dynamic recognition of targets from multiple time series can improve the accuracy of anomaly detection,so time series information is crucial for anomaly detection.In dynamic anomaly detection tasks,the abnormal appearance and motion patterns of objects in the video foreground are key features of anomalous events,so foreground pattern recognition is crucial for anomaly detection.However,the proportion of foreground in video frames usually exhibits some sparsity,and many existing methods learn the patterns of normal events by reconstructing complete video frames.This approach would make the model focus on video background reconstruction while ignoring the learning of foreground pattern features.To alleviate the above problems,this thesis proposes a video anomaly event detection method based on foreground background separation network.This method divides the learning of foreground and background into two subtasks,which guarantees the ability of the model to model normal event patterns.Furthermore,the method utilizes the contour information of optical flow to assist foreground-background separation and to model the motion information of normal events.This thesis conducts experiments on three public datasets and achieves state-of-the-art performance,validating the effectiveness of the foreground-background separation strategy for video anomaly event detection.3.A method for video anomalous behavior detection based on spatiotemporal graph convolutional network and group features.Human abnormal behavior recognition usually relies on longer time series information,which is very common in dynamic anomaly detection,and this behavior has a very serious impact on social security.Therefore,it is of vital significance to study abnormal behavior detection.Existing abnormal behavior detection methods only focus on the movement state within an individual,while ignoring the influence of the interaction state between individuals.In this thesis,we propose an anomalous behavior detection method based on spatiotemporal graph convolutional networks and group features.First,a spatiotemporal graph convolutional network autoencoder is used to extract the fusion features of individual poses and group poses,then a deep clustering algorithm is used to soft-cluster normal behaviors,and finally a Dirichlet process mixture model is used to score individuals anomalies.Experimental results show that the proposed method outperforms the existing video abnormal behavior detection methods in AUC values on two public datasets.更多还原