A comparative evaluation of anomaly detection methods for surveillance videos

The data generated by the camera needs to be analyzed for various reasons. The most important of these is for security reasons. Undesirable and rare events that occurs in videos are considered an anomaly. In recent years, many methods and approaches have been proposed for the detection of anomalies in videos. Although most of these studies have been done on the basis of accuracy, applicability is also important. While it is possible to achieve near-perfect accuracy with complex architectures, the need for memory and processing power increases as the complexity of the related architectures increases. This situation reduces its applicability in real life. Main goal of this study is to review the relatively simple autoencoder-based architectures and two approaches that are frequently used in the literature. These approaches are the raw data approach and the optical flow approach. In the raw data approach, video image data directly imported into the training network and anomaly detection is performed with the spatio-temporal features of the images. In the optical flow approach, the optical flow attribute is calculated from the change between two video frames and taken as an input to the architecture. Thus, motion-based anomaly detection is performed. This study seeks to answer the question of which of the two appraches is better, tested on three different datasets and with three different architectures.