A Research Of Unspecified Anomaly Detection And Localization In Surveillance Videos

Unspecified anomaly detection and localization in surveillance videos have attracted broad attention in both academy and industry for that intelligent surveillance system able to detect anomalies in surveillance videos is very importance to public safety.Heavily depending on human labors,most current surveillance systems can only work as evidence afterwards,but fail to detect anomalies in time,especially under crowded scenes.Aside from localizing normal events on time,it is also important to detect and localize anomalies in existing surveillance videos,which helps policy making departments to improve public safety management such as surveillance facilities installation and security deployment.However,with increasing surveillance facilities and continuous surveillance videos generated by surveillance facilities,it is a mission impossible for humans to handle data in such a large scale.Under such situation,techniques able to detect and localize anomalies in surveillance videos are getting more necessary.In spite of the importance of surveillance systems able to detect and localize anomalies in surveillance videos,the complexity of scenes and deceptiveness of abnormal behaviors make anomaly detection in surveillance videos a challenging task.One challenge of this task is inherent in anomaly detection itself.Anomaly detection is an unsupervised learning task,i.e.,only normal data is contained in training videos,but it is required to find anomalies in test set that contains both normal data and anomalous data,which brings challenges for how to represent normal data and model ‘normality'.Another challenge is caused by surveillance videos.In computer vision community,many image data driven tasks have achieved important breakthroughs.However,tasks based on video data remain challenging in both academy and industry,which mainly caused by incapability to handle temporal information and characteristics.Besides,low-resolution,low-quality properties of surveillance videos also bring challenges for tasks based on them.We will explain such challenges in detail in the following chapters.To handle such challenges,we propose an anomaly detection algorithm called 2streamVAE/GAN in this work,by embedding VAE/GAN in two-stream architectures.Taking both spatial and temporal information into consideration,‘normality' can be captured and anomaly detection can be achieved in our model.To be more specific,we use VAE/GAN to model normal videos in training set,i.e.,to describe and represent ‘normality' with VAE/GAN,then we model normal appearance and temporal information by combining two-stream architectures,so that we can train a model that describes ‘normality'.Based on anomaly detection rules and pre-trained model that defines ‘normality',abnormal events can be automatically detected and localized.We validate our algorithm on two public datasets.Moreover,we design an intelligent surveillance system that is able to detect and localize anomalies in surveillance videos,which can work under both online scenarios,i.e.,to detect anomalies on time,and offline scenarios,i.e.,to detect anomalies in surveillance videos saved in local hard drive or local file system.We will show flow chart,architecture,functions and user interface of our system.