Bayesian Nonparametric Approaches to Abnormality D

Authors: Vu Nguyen Dinh Phung DucSon Pham Svetha Venkatesh

Publish Date: 2015/04/07

Volume: 2, Issue: 1, Pages: 21-41

PDF Link

Abstract

In data science anomaly detection is the process of identifying the items events or observations which do not conform to expected patterns in a dataset As widely acknowledged in the computer vision community and security management discovering suspicious events is the key issue for abnormal detection in video surveillance The important steps in identifying such events include stream data segmentation and hidden patterns discovery However the crucial challenge in stream data segmentation and hidden patterns discovery are the number of coherent segments in surveillance stream and the number of traffic patterns are unknown and hard to specify Therefore in this paper we revisit the abnormality detection problem through the lens of Bayesian nonparametric BNP and develop a novel usage of BNP methods for this problem In particular we employ the Infinite Hidden Markov Model and Bayesian Nonparametric Factor Analysis for stream data segmentation and pattern discovery In addition we introduce an interactive system allowing users to inspect and browse suspicious eventsIn data science anomaly detection is the process of identifying items events or observations which do not conform to expected patterns or other items in a dataset Typically the anomalous items are existing in some kind of specific problem such as bank frauds medical problems or finding errors in text There are two major categories of abnormal detection namely unsupervised abnormal detection and supervised abnormal detection The former detects anomalies in an unlabeled test data set under the assumption that the majority of the instances in the data set are normal by looking for instances that seem to fit least to the remainder of the data set The latter requires a data set that has been labeled as ‘normal’ and ‘abnormal’ and involves training a classifier eg Support Vector Machine 1 Logistic Regression 2In this paper we consider specifically the problem of unsupervised abnormality detection in video surveillance As widely acknowledged in the computer vision community and security management discovering suspicions and irregularities of events in a video sequence is the key issue for abnormal detection in video surveillance 3 4 5 6 7 The important steps in identifying such events include stream data segmentation and hidden patterns discovery However the crucial challenge in stream data segmentation and hidden patterns discovery are the number of coherent segments in surveillance stream and the number of traffic patterns are unknown and hard to specifyThe theory of Bayesian nonparametric BNP 8 9 10 11 12 13 holds a promise to address these challenges As such BNP can automatically identify the suitable number of cluster from the data Therefore in this paper we revisit the abnormality detection problem through the lens of BNP and develop a novel usage of BNP methods for this problem In particular we employ the infinite hidden Markov model 14 and Bayesian nonparametric factor analysis 15The first advantage of our methods is that identifying the unknown number of coherent sections of the video stream would result in better detection performance Each coherent section of motion eg traffic movements at night time and day time would contain different types of abnormality Unlike traditional abnormality detection methods which typically build upon a unified model across data stream The second benefit of our system is an interface allowing users to interactively examine rare events in an intuitive manner Because the abnormal events detected by algorithms and what is considered anomalous by users may be inconsistent the proposed interface would greatly be beneficialTo this end we make two major contributions to abnormal detection in video surveillance 1 proposing to use the Infinite Hidden Markov Model for stream data segmentation and 2 introducing the Bayesian nonparametric Factor Analysisbased interactive system allowing users to inspect and browse suspiciously abnormal eventsThis paper is organized as follows We present an overview on abnormality detection in video surveillance and the need of segmenting the data and interaction in Sect 2 In Sect 3 we describes our contribution on Bayesian nonparametric data stream segmentation for abnormal detection Section 4 illustrates our introduced browsing system for abnormal detection The experiment is provided in Sect 5 Finally we present a summary of the paper with some concluding remarks in Sect 6Ideally abnormality detection algorithms should report only events that require intervention—however this is impossible to achieve with the current stateofart A large semantic gap exists between what is perceived as abnormal and what are computationally realizable outlier events An alternative framework in which the algorithm reports a fraction 1  of rarest events to a human operator for potential intervention 4 has been successful commercially icetanacom By retaining humans in the loop whilst drastically reducing the footage that needs scrutiny the framework provides a practical recourse to machineassisted video surveillance A typical mediumsized city council has to handle hundreds of cameras simultaneously and it is imperative that the computational cost is low This is achieved via an efficient algorithm based on PCA analysis of the video feature data Motionbased features are computed within a fixed duration video clip typically 10–30 s PCA analysis is performed on the training data set to obtain the residual subspace and the threshold corresponding to a desired false alarm rate During testing if the projection of the test vector in the residual subspace exceeds the computed threshold the event is reported to the operator Since the algorithm is based on PCA it is important that the training data is coherent so as to have most of the energy concentrated within a low dimensional principal subspace In this case most normal events remain within the principal space upon projection and the residual subspace retains the fidelity for detecting subtle but rare events However for typical outdoor surveillance the feature vectors generally exhibit different modes depending on the time of day climatic variations etc If we try to fit all these incoherent modes into a single model we reduce the sensitivity of detection If we construct one principle subspace for a 24 h period we are likely to miss events at night because nights have very different motion profiles to that of the daytime

Keywords: