Research On Video Abnormal Event Detection In Complex Scenes Based On Mid-level Semantic Representation

Video abnormal event detection is one of the key tasks in intelligence video surveillance,which aims to identify events caused monitoring targets(human or vehicle,etc.)in intelligence video surveillance that significantly deviated from the expected event model.Due to the increasing attentions drawn in public security management,this problem has recently become a hot topic in computer vision.The key to video abnormal event detection is to obtain normal or abnormal high-level semantic interpretations of video events under specific context by analyzing the original video data.Although low-level feature descriptors are fundamental for abnormal event detection,the descriptive power is still limited and a large semantic gap exists between it and high-level semantic interpretation.In order to reduce the semantic gap and to improve the description of event as well as the performance of the detection system,it is necessary to model the event described by low-level feature descriptors as the more compact,robust,expressive and distinctive mid-level semantic representation.Aiming at the difficulties in video abnormal event detection,such as clutter and crowded scenes,the diversity of the causes of abnormality,the complexity of internal structural in event,the uncertainty and contextual dependency of abnormal event,and the scene dependency of low-level features,in this dissertation,we propose an effective mid-level semantic model for improving the description of event,and we attempt to design a universal abnormal event detection system based on representation learning.Our main contributions of this dissertation are made as follows:1.A novel spatio-temporal cuboid based Gradient-Central-Moments(GCM)descriptor has been proposed.The GCM descriptor is computational efficient,expressive and distinctive,which is used as the basic cues and it can account both the motion and appearance cues in the event.It does not rely on tracking or object detection for features extraction in crowded scenes.Besides,it can be both useful for event description and abnormal event detection in complex scenes.2.A Bag-of-words(BoW)representation based abnormal event detection system has been proposed.Compared with the GCM descriptor,the BoW representation is more robust,expressive and distinctive.Experimental results demonstrate the proposed method can effectively detect various types of abnormal events,and comparable results are gained to state-of-the-art methods.3.A novel Bag-of-structural-context-word(BoSCW)representation based abnormal event detection system has been proposed for utilizing the structural context of event,which is another important aspect in abnormal event detection.The proposed BoSCW model not only inherits the advantages of BoW model,but also overcomes the limitation in the Bag-of-words(BoW)model that is incapable to the model the internal structural contexts of events.Experimental results on different types of datasets suggest that compared with the BoW model-based methods,the performance of proposed method is highly improved.4.A novel Bag-of-atomic-features(BoAF)model based abnormal event detection system has been proposed.Different from the traditional bag-of-features model that assigns only a single visual word in the codebook to a local feature,BoAF model uses a few more atoms in an overcomplete dictionary to approximate local features in a linear combination manner.BoAF model can effectively address the uncertainty problem,since it can model any types of events with low approximation error for any type of local features appeared in the training process or not.Both qualitative and quantitative analysis of experimental results on different datasets show that the better performance is achieved by using BoAF model.5.A novel Hybrid Information(HI)model based abnormal event detection system has been proposed.Considering the contextual dependency of abnormal event,the HI representation of event can account for both the internal information and the external relationship of event.Experimental results show that the performance is better than the method which is only used internal information,and it is better than the state-of-the-art methods.6.A novel representation learning based abnormal event detection system has been proposed.A Deep Incremental Slow Feature Analysis(D-IncSFA)network has been constructed,which can automatically learn the useful multi-sacle semantic representations for abnormal event detection.The proposed method can work well on different types of scenes without depending on any types of handcrafted features.