A Brief Survey on Contemporary Methods for Anomaly Detection in Videos

Computer Vision Applications in Intelligent Transportation Systems: A Survey

“…Unlike the usual video-based action- or event-recognition problems, in which each class is properly identified and labeled, anomaly detection problems are based on learning only the normal data distribution and considering anything that occurs outside this distribution to be an anomaly. For this reason, the video anomaly detection problem can be considered as a one-class problem in which all other classes are unknown [ 268 ].…”

“…The USCD [ 269 ], UMN [ 270 ], and UCF crime datasets [ 271 ] are some of the publicly available datasets used in anomaly detection research. However, when using these datasets, it is difficult to determine whether a network needs to focus on learning motion patterns, object interactions, or something else in order to successfully generalize for an anomaly detection system [ 268 ].…”

“…A brief survey of the contemporary methods developed between 2015 and 2018 for anomaly detection in videos is presented in [ 268 ], which classifies these methods according to their network structures and the datasets used. In anomaly detection using video surveillance cameras, DL-based methods have achieved high performance under harsh environmental conditions [ 272 , 273 , 274 , 275 , 276 ].…”

“…In a relatively similar network proposed in [ 279 ], a Convolution3D-based approach incorporating LSTM was used to extract landmarks from videos. These extracted landmarks were then used to check whether anomalies were present, assuming that the videos contained anomalies [ 268 ].…”

“…Authors utilized CNN for feature extraction in [ 65 , 66 , 67 ]. Following recent advances in ML, several studies have experimented with the use of CNNs, Conv3D, LSTMs, and similar architectures in the field of video anomaly detection [ 268 ].…”