Multi-time Scale Features for Anomaly Detection from Surveillance Videos

As the importance of surveillance in the world increases, the importance of surveillance cameras has increased. It has also increased the number of surveillance cameras installed. On the other hand, as the number of surveillance cameras increases, detecting abnormal events through monitoring requires higher labor intensity. The need for machine resources to detect abnormal events in stored videos has been expanded. Simultaneously, due to the rapid development of the computer vision field, a method has been developed in which the machine can detect abnormal events. To detect abnormal events, the importance of visual features that the machine can recognize has also increased. Previous methods used deep-learned visual features to detect anomalies, while feature extraction of visual features uses only a single stride and a single segment for the time scale. That is, abnormal events are detected by considering only a single scale. These features are spatiotemporal, but there is a possibility that the accuracy can be increased through various strides and segments. Therefore, we propose to consider various strides and various segments, i.e., multi-time scales. We evaluate performance with small datasets reconstructed from existing datasets.