Research On Algorithms For Abnormal Event Detection In Videos

With the propulsion of many projects such as “safe city” and “safe campus”,intelligent video surveillance has become one of the most import techniques in the field of public security.Abnormal event detection is the major route to improve the intellectualized degree of video surveillance equipments.Abnormal event detection aims to automatically analyze video sequences using machine learning and computer version techniques,under the condition of no manual intervention.When suspicious objects appear,the system can automatically detect and analyze them,then send results to the monitoring personnels.The development of this technique can effectively solve the problem of heavily dependence on human resources,and reduce the requirement for storage space.Abnormal event detection is an important and challenging research topic in the field of computer vision.Research on the abnormal event detection algorithms has important academic value and application prospect,and possesses important meanings on national security,public security,and in-home assistance.In recent years,with the progress of video analysis and pattern recognition technologies,the development of abnormal event detection has made great strides.However,there are still some problems: 1)It is inadequate for the research on studying the structures of video events;2)Existing algorithms ignore the study on the composition of abnormal video events;3)Features of video events are not good enough.In order to solve these problems,this paper carries out studies on abnormal event detection algorithms from three aspects.The main research contents and novelties are summarised as follows:1)Structured dictionary learning for abnormal event detection.Most existing algorithms treat video events as individuals,ignoring the structural information of video events.In this case,this paper introduces a new concept named reference event,which describes the typical event type of normal video events.In addition,we propose a structural smoothness regularization,which transmits the structural relationships in the original video sequences and feature space to the space of sparse coding coefficients.Compared with the original dictionary learning based algorithm which does not consider the structural information,the proposed algorithm improves the average detection accuracy by 3.9% on the Avenue dataset.2)Statistical hypothesis detector for abnormal event detection.Traditional algorithms only consider the composition of normal video events,ignoring the information of abnormal events.Abnormal events are identified as samples dissimilar with normal ones.As far as we know,this paper is the first to consider the composition of abnormal events.Abnormal events are identified as samples containing abnormal event patterns,while possessing high abnormality detector scores.Moreover,most existing works assume that the noise is independently sampled from a Gaussian distribution.But in fact,the noise is always sampled from a combination of a variety of distributions.For this reason,this paper models the noise with a mixture of Gaussian,which is proved to be a universal approximation of any continuous density function and is more suitable for real world applications.Meanwhile,according to the anomaly detection results,this paper discovers video events dissimilar with training ones,and uses them to update the normal event patterns.This paper employs three public datasets to verify the superiority of the proposed algorithm,as well as related aspects.3)Deep representation for abnormal event detection.So far,most abnormal event detection algorithms are developed based on hand-crafted features.Generally,designing an effective descriptor is time-consuming and difficult.Moreover,it is hard to decide which kind of feature is suitable for a specific situation.Therefore,this paper proposes a deep representation based algorithm for abnormal event detection.In this method,appearance,texture,and short-term motion features are automatically learned and fused with stacked denoising autoencoders.Due to the time continuous characters of video events,this paper designs a long short-term memory based prediction model to extract the long temporal dependencies of video events.Based on the spatial contextual information,this paper proposes an adaptively weighted manifold ranking algorithm,which considers the relationships among video events while preserving the temporal anomaly scores.This method excludes most false alarms,and improves the discrimination between abnormal and normal events.Experimental results demonstrate the feasibilities of deep features on the video event representation and the recurrent neural network on modeling long temporal video events,respectively.