Research On Semantic Topic Model Based Human Abnormal Behaviour Recognition

Intelligent Video Surveillance (IVS) is a kind of technology to achieve automaticvideo analysis with computer techniques. As an effective means of defense and security,IVS systems are being more and more popular. The analysis and recognition ofpedestrian’s abnormal behavior in the video sequence, a research objective of IVS, hasgradually drawn the attention in the field of IVS. The IVS system based on the analysisof abnormal behavior can not only ignore a large number of useless information, whichguarantees the high efficiency in the security protection, but also save a lot of humanand material resources, which brings great economic benefits to the whole society. Inaddition, it is also able to achieve real-time alarming to eliminate the lags in traditionmonitoring systems.From the perspective of cognitive psychology，the work has been done with thehierarchical structure, following the routine of “Bag of Motion-term–SemanticBehavior Model–Behavior Classification”. The main key point lies in how to train thecomputer to understand the semantic content of behavior category, and recognizesimilarity and difference among behavior categories. The following proposed algorithmand methods are carried out in practice application. The main contributions of this thesisare summarized as follows:1) A number of clustering techniques based on local word-statistics of a video havebeen proposed recently. However, these approaches only captured the content of a videosequence and ignored its order. However, generally behaviors are not fully defined bytheir action-content alone: there are preferred or typical action-orderings. In this work,we addressed these issues by proposing the usage of Hidden Markov Topic Model,HMTM to classify action instances of behavior into states and topics, constructing amore discriminative feature space based on the context-dependent labels, and resultingin potentially better behavior-class discovery and classification. The key advantages ofHMTM over previous approaches are: it is based on constructing a compositegenerative behavior model that scales well with the complexity of behavior and isrobust to errors in behavior representation.2) A number of unsupervised learning of behavior models clustering techniquesbased on local word-statistics of a video have been proposed recently. These approachsusing a "bag-of-words" representation can be categorized into two diffierent types,co-clustering algorithms and topic models. Co-clustering algorithms, applying theclustering results on words as a low dimensional representation of video sequences, can achieve a more accurate clustering for sequences. Topic models, on the other hand, are aclass of statistical models in which the semantic properties of words and sequences areexpressed in terms of probabilistic topics. To solve these problems, we construct anexplicit generative model Co-Clustering Topic Model，CCTM which builds on thestrength of existing co-clustering algorithm and topic models, but crucially is able toovercome their drawbacks on accuracy, robustness and computational efficiency.3) A number of clustering techniques based on visual word-statistics of a videohave been proposed recently. These models represent word co-occurrence, completelyignoring temporal order information. But generally behavior patterns are not fullydefined by their word-content alone; however, there are preferred or typical wordorderings. To this end, a few models have been introduced in which consecutive words(or latent topics) are profiled by Markovian relations. Specifically, hidden topic Markovmodel (HTMM) is similar to the latent Dirichlet allocation (LDA) model in tyingtogether parameters of different documents via a hierarchical generative model, butunlike the LDA model, it does not assume documents are “bags of words”. Rather, itassumes that the topic of words in a document form a Markov chain. Nevertheless,modeling the temporal order of visual words explicitly is risky, because noise in theword representation can easily propagate through the model. In this work, we firstdivide a video sequence into meaningful single-topic segments, then assume segmentsare bag of words, and finally introduce a Markov chain to model segment-topicdynamics. This gains strength in representing temporal information, mean while, it isrobust to noise due to its bag of words modeling of visual words.4) Online anomaly detection using a runtime accumulative anomaly measure andnormal behavior recognition using an online likelihood ratio test (LRT) method. Aruntime accumulative measure is introduced to determine an unseen normal or abnormalbehavior pattern. The behavior pattern is then recognized as one of the normal behaviorcategories using an online LRT method which makes the decision until sufficient visualfeatures have become available. This is to overcome any ambiguity among differentbehavior categories observed online due to insufficient visual evidence at a given timeinstance. By doing so, robust behavior recognition and anomaly detection are ensured assoon as possible, unlike in previous work, which require completed behavior patterns tobe observed. Our online LRT-based behavior recognition approach also has advantagesover previous maximum likelihood (ML) based methods. An ML-based approach makesa forced decision on behavior recognition without considering the reliability andsufficiency of the visual evidence. Consequently, it can be error prone. 5) In recent years, One-class SVM has received considerable attention because oftheir superior performance in anomaly detection. However, most algorithms for SVMproblem require large number of memory to store the kernel matrix, or repeated accessto the training samples, making them unsuitable for anomaly detection. Some onlinealgorithms were proposed for SVM. An important virtue of online algorithms is thatthey can incorporate additional training data, when new sample is available, withoutre-training from scratch. But the methods have to face the the problem of choosing thebest training parameters for optimal clustering. To this end, we propose an alternative,online discriminative anomalous activity detection using one-class support vectormachine (One-class SVM) with PLSA-based Fisher kernel.