Robust Anomaly Detection in CCTV Surveillance

Abstract

Given the vast amount of publicly available CCTV surveillance and the capabilities of modern computer vision algorithms, the task of automatic anomaly detection is due to be solved in the near future. A solution that is competent over the large problem domain requires a certain level of sophistication such that it can replicate the contextual understanding of a human monitor. It is hypothesised that a single approach to anomaly detection can not be expected to perform both low-level and high-level monitoring of video frames which is required for robust anomaly detection. This paper proposes a solution to the anomaly detection problem in the form of a consensus framework that combines inputs from three sources to provide a final verdict on the perceived degree of anomaly contained in a video. The first approach, later introduced as the base model, is an implementation of previous work in anomaly detection that is specifically chosen for its emphasis on the learning of high-level context. The second and third are novel anomaly detection heuristics that operate on a per-frame basis i.e., with no regard for high-level context. The paper concludes with an evaluation and analysis of the three approaches and a discussion of the merit of a consensus framework. A final AUC of 0.7156 is achieved on the UCF Crime dataset; however, this result is not attributable to the consensus framework.