Abstract:Video anomaly detection is challenging due to the diversity of abnormal events, which turns unsupervised learning the promising solution in recent endeavors. In a framework as such, the existing works hold with such an assumption that anomalies cannot be reconstructed or predicated from historical data as accurately as normal ones, so the reconstruction or prediction error can act as an indicator of anomalies. In this study, we propose to discriminate anomalies from normal ones by fusing both appearance and motion in a frame prediction framework, where the uniqueness lies in that we embed optical flows into the frame prediction framework as the clue to direct the transformation from the input to the predicted frame, making appearance-motion fusion quite natural without any extra effort to align them. Then, the error of predicting the next frame based on the concatenation of the appearance and the associated motion of the present frame can indicate the anomaly score. Notably, we employ one frame-based optical flow computation instead of the traditional ones over two continuous frames. The goal is to make the optical flows subject to the whole training data such that anomalies deviating remarkably from the training data will result in highly distorted optical flows and relatively high prediction error correspondingly, which is not promised by the traditional optical flows rooting in the differential of two continuous frames. In summary, we extend the appearance-motion correspondence learning to motion-guided prediction tying the appearances of two consecutive frames. We also introduce a margin loss to enhance the learning of frame prediction. Experiments on widely accepted benchmarks demonstrate the state-of-the-art performance of our approach.

Video Anomaly Detection Based on Mixed Statistic Feature

Anomaly Candidate Identification and Starting Time Estimation of Vehicles from Traffic Videos.

Spatial-based Bayesian Hidden Markov Models with Dirichlet Mixtures for Video Anomaly Detection

Real-time Anomaly Detection with HMOF Feature

Video anomaly detection based on scene classification

Video anomaly detection and localization via Gaussian Mixture Fully Convolutional Variational Autoencoder

Abnormal Behavior Detection Based on Regional Optical Flow

Video Anomaly Detection By The Duality Of Normality-Granted Optical Flow

Video abnormal event detection based on fast sparse coding and surprise computation

Video Anomaly Detection Based on Spatio-Temporal Relationships among Objects

Anomaly Detection and Localisation in the Crowd Scenes Using a Block-Based Social Force Model

Video Anomaly Detection Via Successive Image Frame Prediction Leveraging Optical Flows

Video Anomaly Detection and Localization Using Motion-Field Shape Description and Homogeneity Testing

Robust Anomaly Detection Via Fusion of Appearance and Motion Features

Video Anomaly Detection and Localization Based on an Adaptive Intra-Frame Classification Network

ANOMALY DETECTION AND LOCALIZATION: A NOVEL TWO-PHASE FRAMEWORK BASED ON TRAJECTORY-LEVEL CHARACTERISTICS

MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video Anomaly Detection

Multi-Dimensional Optical Flow Embedded Genetic Programming for Anomaly Detection in Crowded Scenes.

Learning Appearance-motion Normality for Video Anomaly Detection.

Bounding Boxes and Probabilistic Graphical Models: Video Anomaly Detection Simplified