Combining Classifiers for Video Based Surveillance Anomaly Detection

Abstract

In this work we aim to improve the performance of automated surveillance video anomaly detection by introducing two methods for splitting up the detection task into subsets of anomaly types for which we are then able to train independent detectors. First we propose to automatically group scenes based on which objects appear in them, allowing us to use a single detector per scene type. Second we split the problem on anomaly semantics, using independent detectors for fire and smoke detection, and traffic anomaly detection, whose outputs are then combined with the output from a generic anomaly detector to make a final detection prediction. Both of these methods allow models to be trained to detect a smaller range of anomalies which is both a simpler task and allows for a more fine grained interpretation of the reasoning behind detections. We also analyse of how much the current methods are able to distinguish an anomaly from its surrounding normal footage, as present evaluation methods fail to measure this important factor. The results show that fire and smoke detection is the easiest task to separate out and detect independently from generic surveillance video anomaly detection, that the current state of the art methods mostly use scenic priors to detect anomalies, and there are many improvements that need to be made for the state of the art dataset.