Learning and transfer of feature extractors for automatic anomaly detection in surveillance videos

Abstract

Automatic video surveillance is becoming a topic of great importance in the current world. Surveillance cameras in private and public spaces greatly outnumber the humans available for performing the observation task. This hinders the effectiveness of the cameras since the footage is often used much after the event has occurred, rather than allowing for quick corrective action based on real-time detection of events. However, the task of devising robust automatic surveillance systems is a rather difficult one. This high degree of difficulty is associated with the problem of building models able to understand human semantics. Humans have the innate ability to observe an ongoing event and judge its implications, which then leads to decision making. Simulating this understanding in a machine, even to the slightest degree, has become a real challenge in recent research. Computer Vision, Machine Learning, and Deep Learning are fields of study deeply connected to this issue. Together, these fields have recently achieved impressive results across a wide array of vision-related tasks, and provide methods and tools that can be used for the automatic video surveillance problem. In this work, the problem of automatic surveillance is approached from an anomaly detection perspective. It consists of learning a model of normality from videos previously labeled as normal by human observers and then using this model for detecting anomalies. To achieve this goal, the task is divided into two main subtasks: feature extraction and classification. The contributions of this work are mainly related to the feature extraction process, in which two methods based on Deep Learning were proposed. The first method is based on transferring knowledge from a completely unrelated task to video anomaly detection. The idea is to investigate the extent of the generalization capacity of a model, by using it for performing a completely new and unexpected task. The second method is based on learning a feature extractor that extracts compact feature representations from surveillance video datasets. This method was investigated under the hypothesis that creating compact clusters in the feature space may improve the classification performance. Classification is performed by One-Class Support Vector Machines in both methods. Results have shown that the first method had a performance similar to state-of-the-art methods, which leads to the conclusion that the generalization capacity of some Deep Learning models can be extended to different tasks. The results using the second method have corroborated to the compactness hypothesis, in which an increase in classification performance was obtained after introducing compactness. In general, it is possible to conclude that both methods show great promise for enhancing the feature extraction process, and can be valuable contributions towards robust automatic video anomaly detection systems.