Video Anomaly Detection and Localization using 3D SL-HOF Descriptor

Video anomaly detection plays a prominent and challenging role for automated video surveillance. To aim this, we propose a novel framework for local anomaly detection in videos based on 3D Spatially Localized Histogram of Optical Flow (3D SL-HOF) descriptor. The new 3D SL-HOF motion descriptor is capable of capturing global and local motion variations from spatially distributed optical flow map combined with 3D HOF descriptor which efficiently extracts motion velocity and orientation. Each video is described as a set of nonoverlapping spatio-temporal volumes (STVs) and are further partitioned spatially to form 3D local regions. The histogram of optical flow orientation and motion magnitude extracted from motion-rich STVs used as feature descriptor. To reduce computational burden, we compute features for foreground objects. Simple and cost-effective OCSVM classifier is employed to learn normal behaviour during training and detect anomaly from test data. We define Context location to detect abnormal behaviour in an unexpected region. We demonstrate the performance of the proposed method on the benchmarking UCSD Ped1 and Ped2 local anomaly datasets and UMN crowd activity global anomaly dataset. We achieve promising results and compare the performance with state-of-the-art methods.